CN101533475A - Method for extracting feature of shape-adaptive neighborhood based remote sensing image - Google Patents

Abstract

The invention relates to a method for extracting the feature of a shape-adaptive neighborhood based a remote sensing image, which comprises the following steps: selecting an appropriate wave band to form pseudo-color RGB image synthesis and converting the pseudo-color RGB image to an HSI color space; according to the definition of pixel heterogeneity, generating a shape-adaptive neighborhood corresponding to each pixel point of the remote sensing image in the HSI space; extracting the spectral feature, the texture feature and the shape feature of each shape-adaptive neighborhood to generate a plurality of feature patterns; and according to the generated feature patterns; and carrying out feature level data fusion on the spectral feature, the texture feature and the shape feature to form the general feature of the shape-adaptive neighborhood. By using the attention mechanism in human cognitive psychology, the method not only can completely extract the spectral feature, the texture feature and the shape feature of the remote sensing image, but also can better handle the problems of extracting the feature of mixed pixels on a fuzzy edge between the ground and an object.

Description

A kind of remote sensing images feature extracting method based on shape-adaptive neighborhood

Technical field

The present invention relates to the method for a kind of image characteristics extraction in the remote sensing image processing field, it is a kind of remote sensing images feature extracting method based on shape-adaptive neighborhood.

Background technology

(Remote Sensing is that a kind of aeroplane photography that grows up early 1960s is the basic technology of INTEGRATED SIGHT over the ground RS) in remote sensing.Nowadays, the application of remote sensing has been penetrated into the nature and the aspect of social life of human survival, its effect depends on the level of remote sensing image processing, general flow figure such as Fig. 3 of remote sensing image processing, wherein, " feature selecting " in the remote sensing image processing mainly is to select according to demand in spectral band, feature extraction is the committed step in the whole remote sensing image processing process, the validity of the target signature that extracts will directly influence the precision of image classification and Target Recognition, and influence the final effect of remote sensing application.

Traditional feature extracting method can be concluded from three following levels: characteristic layer, method layer and level of abstraction.Wherein, characteristic layer is meant extractible essential characteristic type, comprising: spectral signature, textural characteristics and shape facility; The method layer is to be used for extracting the feature of characteristic layer and the method that adopts, specifically includes spectroscopic analysis methods, texture analysis method and shape analysis method; Level of abstraction is the concrete object of using of the method in the method for expressing layer then, i.e. process object.Traditional method has four kinds of basic processing objects in level of abstraction: the regular neighborhood of single pixel, pixel, the regular piecemeal of image and the object that is partitioned into.Can only the application of spectral analytical approach to single pixel, thus spectral signature can only be extracted; Spectral analysis and texture analysis can be carried out to the regular neighborhood of pixel and the regular piecemeal of image, spectral signature and textural characteristics can be extracted; The object that is partitioned into is handled, can be carried out spectral analysis, texture analysis and shape analysis, thereby can extract spectral signature, textural characteristics and shape facility.In recent years, such processing unit is used for object-oriented Classifying Method in Remote Sensing Image in vogue.Yet for image block or the object that is partitioned into, the original ownership that can not determine picture dot from various characteristic synthetic analyses such as spectrum, texture and shape of cutting apart makes image block or object will include some wrong picture dots that divide.More particularly the mixed pixel of atural object in smeared out boundary is taken as an integral body in their dividing processing, and the wrong picture dot that divides is difficult to obtain correct correction.

Find out that thus there are two deficiencies in existing feature extracting method: 1) traditional feature extracting method can not be complete extracts spectral signature, textural characteristics and shape facility; 2) object oriented analysis method can not good treatment atural object border mixed pixel classification problem.

Cognitive psychology is the subject of a psychological action process in the research human cognitive process, and according to the research of cognitive psychology, human cognitive process comprises following a series of process: attention, sensation, consciousness, memory, presentation and reasoning.In the visual behaviour process, any scene enters the human eye retina, and it is to take optionally to note coming analysis image earlier; By subsequent processes such as sensation and consciousness image is analyzed again; At last, finish identification to image.And this optionally attention is the luminance difference according to image, forms different stimulations at amphiblestroid different cells and finishes.Be subjected to the inspiration of the visual cognition psychology of this attention mechanism, the level of abstraction of the present invention in feature extracting method proposed a kind of new method, that is: at first adopt its shape-adaptive neighborhood to analyze when analyzing the feature of a pixel, the shape-adaptive neighborhood of a pixel then depends on the color characteristic and the shape facility of its surrounding pixel; Then, again it is carried out spectral signature, textural characteristics and Shape Feature Extraction; At last, use it for image classification or Target Recognition.

The content of invention

The objective of the invention is at the problems referred to above, a kind of spectral signature, textural characteristics and shape facility that can extract remote sensing images is provided, can handles the remote sensing images feature extracting method based on shape-adaptive neighborhood of the fuzzy pixel at atural object edge again preferably.

The technical scheme that the present invention takes is: a kind of remote sensing images feature extracting method based on shape-adaptive neighborhood comprises the steps:

(1) it is synthetic to select appropriate wave band to form false colored RGB image, and is transformed into the HSI color space.

(2) on the HSI space of the image that step (1) obtains, the definition pixel heterogeneity to each pixel, generates its shape-adaptive neighborhood (SAN, Shape Adaptive Neighborhood).

(3) SAN that step (2) is generated extracts spectral signature, and textural characteristics and shape facility form many characteristic patterns (feature map).

(4) characteristic pattern of three kinds of features that step (3) is calculated the SAN of gained carries out the data fusion of feature level, generates the general characteristic of neighborhood.

In the step (1), according to demands of applications, adopt the colored composite diagram of the employed vacation of visual interpretation, it is transformed into the HSI color, the formula of conversion is as follows:

$H=\left\{\begin{array}{cc}\arccos \left\{\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}\right\}& R\≠\mathrm{G\; or\; R}\≠B\\ 2\mathrm{\π}-\arccos \left\{\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}\right\}& B>G\end{array}\right.$

$S=1-\frac{3}{R+G+B}\min (R,G,B)$

$I=\frac{R+G+\mathrm{<figure-callout\; id="3"\; label="B"\; filenames="A200910038437E00152.png"\; state="\{\{state\}\}">B</figure-callout>}}{3}$

In the step (2), the heterogeneity between pixel is defined as:

diff＝ω ₁·H+ω ₂·S+ω ₃·I

Wherein, ω _iBe the weight of three components, ω ₁+ ω ₂+ ω ₃=1. like this, for given threshold value T, if diff＜T, then two pixels are atural object of the same race, otherwise different.

Then, the maximal value W*W of given SAN size according to above heterogeneity, can determine the SAN of each pixel.

In the step (3), 1. calculate the spectral signature of SAN, because selected three wave bands are switched to the HSI space, the color characteristic of SAN can be expressed the spectral signature of SAN, and therefore, heterogeneous definition can be used for expressing spectral signature, that is:

SPE＝ω ₁·H+ω ₂·S+ω ₃·I

2. calculate the textural characteristics of SAN, the variation function (variogram) during the geo-statistic after employing improves is learned is expressed the textural characteristics of SAN, and formula is as follows:

$\mathrm{\&gamma;}\left(H\right)=\frac{1}{N\left(H\right)}\underset{i=1}{\overset{N\left(H\right)}{\mathrm{\&Sigma;}}}{[Z\left(x_i\right)-Z(x_i+H)]}^2$

H=[h wherein ₁, h ₂..., h _n] be the step-length sequence of selected calculating texture autocorrelation characteristic, the variation function value that γ (H) promptly calculates also is a proper vector of describing the textural characteristics of this pixel.

3. calculate the shape facility of SAN, adopt two operators of describing compact shape to describe shape facility: outward appearance is than (R) and form factor (F), and its computing formula is as follows:

$R=\frac{L}{W}$ $F=\frac{{\left|B\right|}^2}{4\mathrm{\&pi;A}}$

Wherein, L and W are respectively the length of minimum outsourcing rectangle of neighborhood and wide, and A is the area in zone, and B is the girth in zone.

In addition, also must define the validity of shape.For remote sensing images, because the change of shape of atural object is various, there is not fixed shape, for for atural objects such as natural forest, residential block, farmland, its distribution has bigger randomness, so shape does not have obvious significance for this class atural object; And for the atural object that linear ground objects such as road, river, playground etc. have regular shape, it is very important that the feature of shape then seems.Therefore, need carry out validity constraint to the shape of neighborhood.At this, adopt vector a: eff=[Re, Fe], respectively above two shape description operators are carried out the setting of validity, thereby the shape facility of SAN can be expressed as:

SHA＝[R，F，Re，Fe]

In step (4), need spectral signature, textural characteristics and the shape facility that step (3) is calculated the SAN of gained be merged, thereby obtain the general characteristic of SAN, its formula is as follows:

SANFeature＝Fusion(SPE，γ(H)，SHA)

Wherein, Fusion is the method for Feature Fusion, and SPE is the spectral signature of SAN, and γ (H) is the textural characteristics of SAN, and SHA is the shape facility of SAN, and SANFeature is the total characteristic of the SAN after merging.

Description of drawings

Fig. 1 is based on the implementing procedure figure of the image characteristics extraction of shape-adaptive neighborhood;

The analytic curve figure of Fig. 2 shape-adaptive neighborhood size;

Fig. 3 is the general flow figure of remote sensing image processing.

Embodiment

The implementing procedure figure of the remote sensing images feature extracting method based on shape-adaptive neighborhood of the present invention in the present embodiment, is that the high-resolution remote sensing image of 5m is an example with the resolution of SPOT-5 as shown in Figure 1, and it comprises four steps:

(1) it is synthetic to select appropriate wave band to form false colored RGB image, and is transformed into the HSI color space.According to demands of applications, to select the colored composite diagram of the employed vacation of visual interpretation, and it is transformed into the HSI color, the formula of conversion is as follows:

$H=\left\{\begin{array}{cc}\arccos \left\{\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}\right\}& R\&NotEqual;\mathrm{G\; or\; R}\&NotEqual;B\\ 2\mathrm{\&pi;}-\arccos \left\{\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}\right\}& B>G\end{array}\right.$

$S=1-\frac{3}{R+G+B}\min (R,G,B)$

$I=\frac{R+G+\mathrm{<figure-callout\; id="3"\; label="B"\; filenames="A200910038437E00152.png"\; state="\{\{state\}\}">B</figure-callout>}}{3}$

(2) on the HSI space of the image that step (1) obtains, the definition pixel heterogeneity to each pixel, generates its shape-adaptive neighborhood (SAN, Shape Adaptive Neighborhood).Heterogeneity between pixel is defined as:

diff＝ω ₁·H+ω ₂·S+ω ₃·I

Wherein, ω _iBe the weight of three components, ω ₁+ ω ₂+ ω ₃=1. like this, for given threshold value T, if diff＜T, then two pixels are atural object of the same race, otherwise different.

Then, the maximal value W*W of given SAN size according to above heterogeneity, can determine the SAN of each pixel.

(3) SAN that step (2) is generated extracts spectral signature, and textural characteristics and shape facility form many characteristic patterns (feature map).

1. calculate the spectral signature of SAN, because selected three wave bands are switched to the HSI space, the color characteristic of SAN can be expressed the spectral signature of SAN, and therefore, heterogeneous definition can be used for expressing spectral signature, that is:

SPE＝ω ₁·H+ω ₂·S+ω ₃·I

2. calculate the textural characteristics of SAN, the variation function (variogram) during the geo-statistic after employing improves is learned is expressed the textural characteristics of SAN, and formula is as follows:

$\mathrm{\&gamma;}\left(H\right)=\frac{1}{N\left(H\right)}\underset{i=1}{\overset{N\left(H\right)}{\mathrm{\&Sigma;}}}{[Z\left(x_i\right)-Z(x_i+H)]}^2$

H=[h wherein ₁, h ₂..., h _n] be the step-length sequence of selected calculating texture autocorrelation characteristic, the variation function value that γ (H) promptly calculates also is a proper vector of describing the textural characteristics of this pixel.

3. calculate the shape facility of SAN, adopt two operators of describing compact shape to describe shape facility: outward appearance is than (R) and form factor (F), and its computing formula is as follows:

$R=\frac{L}{W}$ $F=\frac{{\left|B\right|}^2}{4\mathrm{\&pi;A}}$

Wherein, L and W are respectively the length of minimum outsourcing rectangle of neighborhood and wide, and A is the area in zone, and B is the girth in zone.

In addition, also must define the validity of shape.For remote sensing images, because the change of shape of atural object is various, there is not fixed shape, for for atural objects such as natural forest, residential block, farmland, its distribution has bigger randomness, so shape does not have obvious significance for this class atural object; And for the atural object that linear ground objects such as road, river, playground etc. have regular shape, it is very important that the feature of shape then seems.Therefore, need carry out validity constraint to the shape of neighborhood.At this, adopt vector a: eff=[Re, Fe], respectively above two shape description operators are carried out the setting of validity, thereby the shape facility of SAN can be expressed as:

SHA＝[R，F，Re，Fe]

(4) characteristic pattern of three kinds of features that step (3) is calculated the SAN of gained carries out the data fusion of feature level, generates the general characteristic of neighborhood.Need spectral signature, textural characteristics and the shape facility of SAN be merged, thereby obtain the general characteristic of SAN, its formula is as follows:

SANFeature＝FuSion(SPE，γ(H)，SHA)

Wherein, Fusion is the method for Feature Fusion, and SPE is the spectral signature of SAN, and γ (H) is the textural characteristics of SAN, and SHA is the shape facility of SAN, and SANFeature is the total characteristic of the SAN after merging.

In order to verify validity of the present invention, we utilize remote sensing images feature extracting method provided by the invention to carry out following various tests.Shown in Fig. 2～4, in the test, remote sensing images to a SPOT-5 carry out feature extraction, adopt maximum likelihood to estimate that sorter and ISODATA sorter carry out the classification of feature space then, the weight of H, S and I component is respectively 0.7,0.1 and 0.2 in the heterogeneous definition, heterogeneous threshold value is 0.05, and the shape-adaptive neighborhood maximal value is set to 21*21, and the step-length that variation function is calculated in the texture feature extraction gets 1,2,3,4 and 5 respectively.Then, the size of all SAN is analyzed, and the nicety of grading in different characteristic space is carried out computational analysis.In addition, give the quantitative test of maximal value with the heterogeneous threshold value of the best in optimal self-adaptive field.

1, to the analysis of all SAN sizes in the image.The size of all SAN in the statistical experiment image, as shown in Figure 2, unnecessary 20 of the number of pixels that the SAN above 91% comprises satisfies the requirement of in the subsequent step SAN being carried out texture analysis and shape analysis.

2, adopt maximum likelihood to estimate that sorter carries out the classification in different characteristic space.Adopt maximum likelihood to estimate sorter, utilize the original spectrum feature of trial image respectively, the color characteristic of SAN and the color of SAN and textural characteristics are classified.As Fig. 2, it is six classes that image is divided into: water body, arable land, forest land, settlement place, development area and unfiled.Among the subgraph b, adopt spectral signature, some arable lands are divided into the forest land by mistake; Adopt color characteristic (subgraph c) arable land of then can correctly classifying, yet but some shade mistakes are divided into water body; The fusion (subgraph d) of color characteristic and textural characteristics by SAN is just ploughing and shade is correctly classified.

3, the nicety of grading that adopts different characteristic to classify is analyzed.Adopt supervised classification (maximum likelihood estimation) and unsupervised classification (ISODATA) respectively, original spectrum feature to trial image, color and the textural characteristics of the color characteristic of SAN, SAN textural characteristics and SAN are classified, nicety of grading from table 1 can be seen, adopt color characteristic and the textural characteristics of SAN can improve nicety of grading, and use the fusion feature of color and textural characteristics will make nicety of grading higher.

Table 1

4, the peaked analysis in optimal self-adaptive field.The shape-adaptive neighborhood of a pixel (SAN) is minimum to comprise a pixel, can comprise whole image at most, if but at every turn all whole picture search, then can reduce the efficient of feature extraction greatly, therefore, before the SAN that determines each pixel, need the maximal value (hunting zone) of SAN be limited, in fact, the atural object on the remote sensing images can be greatly to not covering whole figure, so this restriction is rational yet.But different threshold limits may exert an influence to the effect of feature extraction, this experimental analysis under the different SAN size threshold value to the influence of final nicety of grading.As shown in table 2, for the test remote sensing images, the threshold value of best SAN size is that 7*7 is between the 11*11, threshold value greater than 15*15 can make nicety of grading descend, this is to increase the more noise picture dot because bigger hunting zone may be SAN, and this noise pixel is also relevant with pixel heterogeneity threshold value to be discussed below.

Table 2

Maximum size of SAN

Total precision

Kappa

(pixels)
			7*7	0.9023	0.7893
11*11	0.9085	0.7816
			15*15	0.9026	0.7620
21*21	0.8954	0.7463
			31*31	0.8800	0.7252
41*41	0.8718	0.7249

5, the analysis of Zui Jia heterogeneous threshold value.In fact, the threshold value acting in conjunction of pixel heterogeneity threshold value and SAN size and finally determine the SAN of each pixel, the too small heterogeneous threshold value of numerical value can make the not integrity property of thing corresponsively of SAN, excessive threshold value then can make SAN that the pixel outside the atural object is comprised to come in, thereby forms the quality that noise influences feature extraction thereafter.This experiment is calculated by the nicety of grading to feature space, analyzes the influence of different heterogeneous threshold values to feature extraction.As shown in table 3, when heterogeneous threshold value was 0.5, the nicety of grading of remote sensing images and Kappa coefficient reached the highest.

Table 3

Threshold of the heterogeneity	Totalpr ecision	Kappa
			0.01	0.8909	0.7668
0.10	0.9061	0.7805
			0.20	0.9080	0.7722
0.50	0.9131	0.7883
			0.80	0.8968	0.7574

Claims (5)

1, a kind of remote sensing images feature extracting method based on shape-adaptive neighborhood is characterized in that comprising the steps:

(1) it is synthetic to select appropriate wave band to form false colored RGB image, and is transformed into the HSI color space;

(2) on the HSI space of the image that step (1) obtains, the definition pixel heterogeneity to each pixel, generates its shape-adaptive neighborhood;

(3) SAN that step (2) is generated extracts spectral signature, and textural characteristics and shape facility form many characteristic patterns;

(4) characteristic pattern of three kinds of features that step (3) is calculated the SAN of gained carries out the data fusion of feature level, generates the general characteristic of neighborhood.

2, the remote sensing images feature extracting method based on shape-adaptive neighborhood according to claim 1 is characterized in that adopting the colored composite diagram of the employed vacation of visual interpretation in the above-mentioned steps (1), and it is transformed into the HSI color, and the formula of conversion is as follows:

$H=\left\{\begin{array}{cc}\arccos \left\{\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}\right\}& R\&NotEqual;\mathrm{G\; or\; R}\&NotEqual;B\\ 2\mathrm{\&pi;}-\arccos \left\{\frac{(R-G)+(R-B)}{2\sqrt{{(R-G)}^2+(R-B)(G-B)}}\right\}& B>G\end{array}\right.$

$S=1-\frac{3}{R+G+B}\min (R,G,B)$

$I=\frac{R+G+B}{3}$

3, the remote sensing images feature extracting method based on shape-adaptive neighborhood according to claim 1 is characterized in that the heterogeneity between pixel is defined as in the above-mentioned steps (2):

diff＝ω ₁·H+ω ₂·S+ω ₃·I

Wherein, ω _iBe the weight of three components, ω ₁+ ω ₂+ ω ₃=1. like this, for given threshold value T, if diff＜T, then two pixels are atural object of the same race, otherwise different;

Then, the maximal value W*W of given SAN size according to above heterogeneity, can determine the SAN of each pixel.

4, the remote sensing images feature extracting method based on shape-adaptive neighborhood according to claim 1 is characterized in that comprising three aspects in the above-mentioned steps (3),

1. calculate the spectral signature of SAN, because selected three wave bands are switched to the HSI space, the color characteristic of SAN can be expressed the spectral signature of SAN, and therefore, heterogeneous definition can be used for expressing spectral signature, that is:

SPE＝ω ₁·H+ω ₂·S+ω ₃·I

2. calculate the textural characteristics of SAN, the variation function (variogram) during the geo-statistic after employing improves is learned is expressed the textural characteristics of SAN, and formula is as follows:

$\mathrm{\&gamma;}\left(H\right)=\frac{1}{N\left(H\right)}\underset{i=1}{\overset{N\left(H\right)}{\mathrm{\&Sigma;}}}{[Z\left(x_i\right)-Z(x_i+H)]}^2$

H=[h wherein ₁, h ₂..., h _n] be the step-length sequence of selected calculating texture autocorrelation characteristic, the variation function value that γ (H) promptly calculates also is a proper vector of describing the textural characteristics of this pixel;

3. calculate the shape facility of SAN, adopt two operators of describing compact shape to describe shape facility: outward appearance is than (R) and form factor (F), and its computing formula is as follows:

$R=\frac{L}{W}$ $F=\frac{{\left|B\right|}^2}{4\mathrm{\&pi;A}}$

Wherein, L and W are respectively the length of minimum outsourcing rectangle of neighborhood and wide, and A is the area in zone, and B is the girth in zone;

In addition, also must define the validity of shape.For remote sensing images, because the change of shape of atural object is various, there is not fixed shape, for for atural objects such as natural forest, residential block, farmland, its distribution has bigger randomness, so shape does not have obvious significance for this class atural object; And for the atural object that linear ground objects such as road, river, playground etc. have regular shape, it is very important that the feature of shape then seems.Therefore, need carry out validity constraint to the shape of neighborhood; At this, adopt vector a: eff=[Re, Fe], respectively above two shape description operators are carried out the setting of validity, thereby the shape facility of SAN can be expressed as:

SHA＝[R，F，Re，Fe]。

5, the remote sensing images feature extracting method based on shape-adaptive neighborhood according to claim 1, it is characterized in that in the above-mentioned steps (4), need spectral signature, textural characteristics and the shape facility that step (3) is calculated the SAN of gained be merged, thereby obtain the general characteristic of SAN, its formula is as follows:

SANFeature＝Fusion(SPE，γ(H)，SHA)

Wherein, Fusion is the method for Feature Fusion, and SPE is the spectral signature of SAN, and γ (H) is the textural characteristics of SAN, and SHA is the shape facility of SAN, and SANFeature is the total characteristic of the SAN after merging.

Images