CN104794495A - Large-format remote-sensing image region classifying method based on straight line statistical characteristics - Google Patents

Abstract

The invention relates to a large-format remote-sensing image region classifying method based on straight line statistical characteristics. The large-format remote-sensing image region classifying method based on the straight line statistical characteristics comprises the following steps of (1) uniform grid division, (2) extraction of a straight line, (3) extraction of a straight line supporting region, (4) ellipse fitting, (5) redefining of the straight line, (6) extraction of length attributes of the straight line, (7) calculation of the statistical characteristic of the straight line and (8) classification. Different types of remote-sensing image regions have different straight line statistical characteristics, types of wastelands, countries, suburbs, urban agglomerative regions and the like in a large-format remote-sensing image can be divided accurately, the problem that buildings are searched in the remote-sensing image blindly is solved, accuracy and efficiency of detection of the buildings are improved, and a complete automation effect is achieved.

Description

A kind of large format remote sensing image territorial classification method based on straight line statistical nature

Technical field

The present invention relates to a kind of remote sensing image process field, is a kind of large format remote sensing image territorial classification method based on straight line statistical nature specifically.

Background technology

Traditional remote sensing image territorial classification method based on pixel is only in " image procossing " stage in Image Engineering, can not meet the requirement of modern Remote Sensing Technical development.For high-resolution remote sensing image, single pixel generally can not reflect real geographic object, and the topological relation between pixel is also very limited.Owing to have ignored the space characteristics such as texture, context and shape based on the disposal route of pixel, even if soft sorter, sub-pixed mapping sorting technique and spectrum solution technology of mixing is introduced wherein, only based on image spectrum information classification results also difficulty have and significantly improve.When classifying to high-resolution remote sensing image, the method based on pixel can cause even more serious " spiced salt effect ", thus affects the accuracy of classification results.Although the sensor information extracting method precision of traditional visual interpretation is higher, need to drop into a large amount of manpowers and time, oneself can not the processing requirements of satisfying magnanimity data message.Current, from high-resolution remote sensing image, the means of extracting geographic information are also relatively backward, and comparatively distinct issues remain " data magnanimity, information are not enough, knowledge is difficult asks ".In addition, the people be in information society more and more pay attention to the ageing of information, and this effectively facilitates Remote Sensing Data Processing method towards future development that is semi-automatic, robotization.Therefore, efficient, intelligentized remote sensing image territorial classification certainly will become study hotspot that is current and even remote sensing information process from now on.

Summary of the invention

The invention provides a kind of large format remote sensing image territorial classification method based on straight line statistical nature, buildings in current remote sensing image can be overcome and extract the low problem of accuracy rate, make full use of dissimilar remote sensing image region and there is different straight line statistical natures, can distinguish exactly comprise in large format remote sensing image wasteland, rural area, the type such as outskirts of a town and High-Density Urban Area, without the need to manual intervention, automaticity is high.

The technical scheme adopted for realizing target of the present invention is: method comprises the following steps:

Step 1: divide by uniform grid input remote sensing image image1, with the subimage I after dividing ₁, I ₂..., I _nas minimal processing unit, n is the sum of the subimage after dividing;

Step 2: to subimage I ₁, I ₂..., I _nstraight Line Extraction is used to extract straight line respectively;

Step 3: respectively to subimage I ₁, I ₂..., I _nin meet syntople straight line merge into a straight line support region;

Step 4: straight line support region is fitted to ellipse with Fourier descriptor;

Step 5: extract oval major axis, and be newly defined as straight line;

Step 6: the length attribute extracting straight line;

Step 7: the statistical nature of calculated line, comprises average and the entropy of straight length;

Step 8: utilize the statistical nature of sorter to the straight line in step 7 to classify.

Described Straight Line Extraction adopts the method based on gradient, the gradient G in x and y direction _x(x, y) and G _y(x, y) utilizes following two formulae discovery respectively:

G _x(x,y)＝{e(x)h(y)*f(x,y)}

G _y(x,y)＝{e(y)h(x)*f(x,y)}

The gradient calculation formula of pixel (x, y) is:

$G(x,y)=\arctan \frac{G_x(x,y)}{G_y(x,y)}$

Wherein, e (k) is boundary filter, and h (k) is mapped filter, and the value of k is x or y, and corresponding formula is as follows:

$e\left(k\right)=\frac{e^{-\mathrm{\αk}}}{\mathrm{\α}}[-\frac{\cos (\mathrm{\α\βk}+\mathrm{\π}/2)-\mathrm{\β}\sin (\mathrm{\α\βk}+\mathrm{\π}/2)}{1+{\mathrm{\β}}^2}]$

h(k)＝e ^-αkcos(αβk+π/2)

Wherein α is scale parameter, and β is image resolution parameter.

The defining method of described syntople is: choose a threshold value T (threshold value T is set as 5), if the bee-line between two straight lines is less than T, then judges that these two is directly adjacent, otherwise is non-adjacent.

Described with Fourier descriptor, straight line support region fitted to oval method and is: utilize Fourier expansion and Euler's formula, change is carried out to multiply periodic function and obtains following Fourier coefficient:

$a_n=\frac{1}{T}[\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}x\left(k\right)\cos \left(\frac{2\mathrm{\πnk}}{T}\right)+\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}y\left(k\right)\sin \left(\frac{2\mathrm{\πnk}}{T}\right)]$

${\mathrm{\β}}_n=\frac{1}{T}[\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}y\left(k\right)\cos \left(\frac{2\mathrm{\πnk}}{T}\right)+\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}x\left(k\right)\sin \left(\frac{2\mathrm{\πnk}}{T}\right)]$

Wherein x (k) and y (k) represents real part and the imaginary part of the borderline point of straight line support region respectively.

By extracting three groups of coefficients: (α _-1, β _-1), (α ₀, β ₀) and (α ₁, β ₁), three oval parameters can be obtained:

Center: c _xy=(α ₀, β ₀)

Length: $L=2[\sqrt{{\mathrm{\α}}_1^2+{\mathrm{\β}}_1^2}+\sqrt{{\mathrm{\α}}_{-1}^2+{\mathrm{\β}}_{-1}^2}]$

Direction: $\mathrm{\θ}=\frac{\arctan \left(\frac{{\mathrm{\β}}_1}{{\mathrm{\α}}_1}\right)+\arctan \left(\frac{{\mathrm{\β}}_{-1}}{{\mathrm{\α}}_{-1}}\right)}{2}$

The computing method of the entropy of the statistical nature cathetus length of described straight line are: build a histogram by 50 bin, the width of each bin is 4 pixels, utilizes following formula to calculate:

$E=-\underset{i=1}{\overset{50}{\mathrm{\Σ}}}h\left(i\right){\log }_2\left(h\left(i\right)\right)$

Described sorter adopts Pa Ersen window sorter.

The invention has the beneficial effects as follows: solve the problem of searching for buildings in remote sensing image blindly, be conducive to the accuracy and efficiency improving buildings detection, reach full automatic effect.

Accompanying drawing explanation

Fig. 1 is overall process flow figure of the present invention.

Embodiment

The specific embodiment of the present invention is described in detail below in conjunction with accompanying drawing.

In step 101, the pending remote sensing image of input is Quick bird high spatial resolution remote sense image image1, is of a size of 8000 × 8000, and carries out the pre-service such as radiant correction and geometry correction.

In step 102, remote sensing image image1 is divided into the subimage of 400 400 × 400 by uniform grid.

In step 103, respectively to each 400 × 400 subimage, adopt the method compute gradient based on gradient, the gradient G in x and y direction _x(x, y) and G _y(x, y) utilizes following two formulae discovery respectively:

G _x(x,y)＝{e(x)h(y)*f(x,y)}

G _y(x,y)＝{e(y)h(x)*f(x,y)}

The gradient calculation formula of pixel (x, y) is:

$G(x,y)=\arctan \frac{G_x(x,y)}{G_y(x,y)}$

Wherein, e (k) is boundary filter, and h (k) is mapped filter, and the value of k is x or y, and corresponding formula is as follows:

$e\left(k\right)=\frac{e^{-\mathrm{\αk}}}{\mathrm{\α}}[-\frac{\cos (\mathrm{\α\βk}+\mathrm{\π}/2)-\mathrm{\β}\sin (\mathrm{\α\βk}+\mathrm{\π}/2)}{1+{\mathrm{\β}}^2}]$

h(k)＝e ^-αkcos(αβk+π/2)

Wherein α is scale parameter, and getting 1.8, β is image resolution parameter, gets 70.

Finally, the pixel that gradient magnitude is less than 10 is got rid of, for restraint speckle.

In step 104, the straight line that the subimage for each 400 × 400 extracts in step 103, selected threshold T=5, calculates the bee-line L between two straight lines _minif, L _minbe less than T, then judge that these two is directly adjacent, otherwise be non-adjacent, the straight line of all of its neighbor is merged into a straight line support region.

In step 105, straight line support region step 104 obtained with Fourier descriptor fits to ellipse: utilize Fourier expansion and Euler's formula, carries out change obtain following Fourier coefficient to multiply periodic function:

$a_n=\frac{1}{T}[\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}x\left(k\right)\cos \left(\frac{2\mathrm{\πnk}}{T}\right)+\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}y\left(k\right)\sin \left(\frac{2\mathrm{\πnk}}{T}\right)]$

${\mathrm{\β}}_n=\frac{1}{T}[\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}y\left(k\right)\cos \left(\frac{2\mathrm{\πnk}}{T}\right)+\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}x\left(k\right)\sin \left(\frac{2\mathrm{\πnk}}{T}\right)]$

Wherein x (k) and y (k) represents real part and the imaginary part of the borderline point of straight line support region respectively.

By extracting three groups of coefficients: (α _-1, β _-1), (α ₀, β ₀) and (α ₁, β ₁), three oval parameters can be obtained:

Center: c _xy=(α ₀, β ₀)

Length: $L=2[\sqrt{{\mathrm{\α}}_1^2+{\mathrm{\β}}_1^2}+\sqrt{{\mathrm{\α}}_{-1}^2+{\mathrm{\β}}_{-1}^2}]$

Direction: $\mathrm{\θ}=\frac{\arctan \left(\frac{{\mathrm{\β}}_1}{{\mathrm{\α}}_1}\right)+\arctan \left(\frac{{\mathrm{\β}}_{-1}}{{\mathrm{\α}}_{-1}}\right)}{2}$

In step 106, its major axis of ellipse step 105 obtained represents, is in line by ellipse fitting.

In step 107, the statistical nature of calculated line, comprises average and the entropy of straight length.

The computing formula of the average of straight length is:

$M=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{L}_i$

Wherein, N is the quantity of last fitting a straight line in subimage, L _ibe the length of i-th straight line

The computing method of the entropy of straight length are: build a histogram by 50 bin, the width of each bin is 4 pixels, utilizes following formula to calculate:

$E=-\underset{i=1}{\overset{50}{\mathrm{\Σ}}}h\left(i\right){\log }_2\left(h\left(i\right)\right)$

In step 108, Pa Ersen window sorter is adopted to classify.

In step 109, output category result: in input remote sensing image image1, classification results pseudo-colours superposed, wherein, represent wasteland by yellow, green expression rural area, blueness represents outskirts of a town, red expression High-Density Urban Area.

Claims (6)

1., based on a large format remote sensing image territorial classification method for straight line statistical nature, it is characterized in that comprising the following steps:

Step 1: divide by uniform grid input remote sensing image image1, with the subimage I after dividing ₁, I ₂..., I _nas minimal processing unit, n is the sum of the subimage after dividing;

Step 2: to subimage I ₁, I ₂..., I _nstraight Line Extraction is used to extract straight line respectively;

Step 3: respectively to subimage I ₁, I ₂..., I _nin meet syntople straight line merge into a straight line support region;

Step 4: straight line support region is fitted to ellipse with Fourier descriptor;

Step 5: extract oval major axis, and be newly defined as straight line;

Step 6: the length attribute extracting straight line;

Step 7: the statistical nature of calculated line, comprises average and the entropy of straight length;

Step 8: utilize the statistical nature of sorter to the straight line in step 7 to classify.

2. a kind of large format remote sensing image territorial classification method based on straight line statistical nature according to claim 1, is characterized in that Straight Line Extraction adopts the method based on gradient, the gradient G in x and y direction _x(x, y) and G _y(x, y) utilizes following two formulae discovery respectively:

G _x(x,y)＝{e(x)h(y)*f(x,y)}

G _y(x,y)＝{e(y)h(x)*f(x,y)}

The gradient calculation formula of pixel (x, y) is:

$G(x,y)=\arctan \frac{G_x(x,y)}{G_y(x,y)}$

Wherein, e (k) is boundary filter, and h (k) is mapped filter, and the value of k is x or y, and corresponding formula is as follows:

$e\left(k\right)=\frac{e^{-\mathrm{ak}}}{\mathrm{\α}}[-\frac{\cos (\mathrm{\α\βk}+\mathrm{\π}/2)-\mathrm{\β}\sin (\mathrm{\α\βk}+\mathrm{\π}/2)}{1+{\mathrm{\β}}^2}]$

h(k)＝e ^-αkcos(αβk+π/2)

Wherein α is scale parameter, and β is image resolution parameter.

3. a kind of large format remote sensing image territorial classification method based on straight line statistical nature according to claim 1, it is characterized in that the defining method of syntople is: choose a threshold value T (threshold value T is set as 5), if the bee-line two between straight line is less than T, then judge that these two is directly adjacent, otherwise be non-adjacent.

4. a kind of large format remote sensing image territorial classification method based on straight line statistical nature according to claim 1, it is characterized in that with Fourier descriptor, straight line support region being fitted to oval method is: utilize Fourier expansion and Euler's formula, change is carried out to multiply periodic function and obtains following Fourier coefficient:

$a_n=\frac{1}{T}[\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}x\left(k\right)\cos \left(\frac{2\mathrm{\πnk}}{T}\right)+\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}y\left(k\right)\sin \left(\frac{2\mathrm{\πnk}}{T}\right)]$

${\mathrm{\β}}_n=\frac{1}{T}[\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}y\left(k\right)\cos \left(\frac{2\mathrm{\πnk}}{T}\right)-\underset{k=0}{\overset{T-1}{\mathrm{\Σ}}}x\left(k\right)\sin \left(\frac{2\mathrm{\πnk}}{T}\right)]$

Wherein x (k) and y (k) represents real part and the imaginary part of the borderline point of straight line support region respectively.

By extracting three groups of coefficients: (α _-1, β _-1), (α ₀, β ₀) and (α ₁, β ₁), three oval parameters can be obtained:

Center: c _xy=(α ₀, β ₀)

Length: $L=2[\sqrt{{\mathrm{\α}}_1^2+{\mathrm{\β}}_1^2}+\sqrt{{\mathrm{\α}}_{-1}^2+{\mathrm{\β}}_{-1}^2}]$

Direction: $\mathrm{\θ}=\frac{\arctan \left(\frac{{\mathrm{\β}}_1}{{\mathrm{\α}}_1}\right)+\arctan \left(\frac{{\mathrm{\β}}_{-1}}{{\mathrm{\α}}_{-1}}\right)}{2}$

5. a kind of large format remote sensing image territorial classification method based on straight line statistical nature according to claim 1, it is characterized in that the computing method of the entropy of the statistical nature cathetus length of straight line are: build a histogram by 50 bin, the width of each bin is 4 pixels, utilizes following formula to calculate:

$E=\underset{i=1}{\overset{50}{\mathrm{\Σ}}}h\left(i\right){\log }_2\left(h\left(i\right)\right)$

6. a kind of large format remote sensing image territorial classification method based on straight line statistical nature according to claim 1, is characterized in that sorter adopts Pa Ersen window sorter.