CN107578446A - A method and device for extracting roads from remote sensing images - Google Patents

Abstract

The present invention relates to a kind of method for extracting remote sensing image road and device, belongs to remote sensing information extractive technique field.The present invention calculates original remote sensing image road probability density first, and the roadway characteristic on original remote sensing image is converted to the feature of road probability density；Then according to feature construction road-center line model of the probable value on road axis higher than the probable value in other positions, and cost function is determined according to road-center line model；The maximum of cost function is finally solved using Dynamic Programming, road axis is used as using the maximum.Complicated and diversified roadway characteristic on original remote sensing image is converted to simple consistent roadway characteristic in road probability distribution graph by the present invention, enable to extract different types of high-resolution remote sensing image road, cost function need not be changed, improves the universality of method.

Description

A method and device for extracting roads from remote sensing images

technical field

The invention relates to a remote sensing image road extraction method and device, belonging to the technical field of remote sensing information extraction.

Background technique

Automatically extracting roads from high-resolution remote sensing images is of great significance for map updating, GIS data acquisition, image matching and target detection, and is the current research focus in the field of remote sensing mapping. High-resolution remote sensing image road extraction can be divided into fully automatic extraction and semi-automatic extraction. The common automatic extraction methods are mainly based on parallel line pairs, based on mathematical morphology and knowledge, and based on window model features. , the existing fully automatic algorithm has poor robustness, and the extraction results require a lot of manual processing, and the effect is not ideal. The semi-automatic extraction using human-computer interaction is currently a more practical choice.

In the semi-automatic road extraction methods of remote sensing images, the active contour model and the template matching method are considered to be two more practical methods. The road extraction algorithm based on dynamic programming is a commonly used active contour model method. It constructs a cost function based on the road features on remote sensing images, and then uses dynamic programming to solve the maximum value of the cost function to extract roads. Armin Gruen proposed a classic road extraction algorithm based on dynamic programming based on the feature that roads on low-resolution remote sensing images are mainly smooth curves with high gray values, but this method is only suitable for low-resolution images , while in the medium and high resolution images, the road is no longer a simple linear feature, but a long strip area with a certain width. Factor Poz et al. modified the cost function on the basis of Gruen, adding the road width information, so that the algorithm can be used for road extraction in medium and high resolution images. However, due to the complex and changeable road characteristics of high-resolution images, traditional algorithms define the cost function directly based on the road grayscale features on the original image, so it is difficult to define a universal cost function, resulting in traditional algorithms that can only extract For simple roads with fixed grayscale features, for other types of roads, the corresponding cost function can only be redefined, which has great limitations in practical applications.

Contents of the invention

The purpose of the present invention is to provide a remote sensing image road extraction method to solve the problem that the current road extraction method can only extract simple roads that match the road model; the present invention also provides a remote sensing image road extraction device.

In order to solve the above-mentioned technical problems, the present invention provides a method for extracting roads from remote sensing images, including six technical solutions, method solution 1, which includes the following steps:

1) Calculate the road probability density of the original remote sensing image, and convert the road features on the original remote sensing image into the features of the road probability density;

2) Construct a road centerline model according to the characteristic that the probability value on the road centerline is higher than that on other positions, and determine the cost function according to the road centerline model;

3) Use dynamic programming to solve the maximum value of the cost function, and use this maximum value as the road centerline.

The invention converts complex and diverse road features on the original remote sensing image into simple and consistent road features on the road probability distribution map, so that it can extract different types of high-resolution remote sensing image roads without modifying the cost function, and improves the generality of the method. fitness.

Method scheme two: on the basis of method scheme one, the road probability density of the remote sensing image in step 1) is determined by using support vector machine and kernel density estimation, and the specific process is as follows:

A. Using support vector machine to classify the remote sensing image, get a series of road sample points;

B. Calculate the probability density of each road sample point using kernel density estimation.

The present invention uses a support vector machine for classification, the required training samples are manually selected from the road feature database, and a small amount of manual participation is used to complete the task of collecting road information with a certain width on the high-resolution remote sensing image, which significantly shortens the mapping cycle. The degree of automatic processing of remote sensing images has been greatly improved.

Method scheme three: on the basis of method scheme two, the probability density obtained in the step B is:

Where x _i is the i-th road sample point obtained, is the probability density estimate at point x, h is the bandwidth of the kernel density estimate, and K(x) is the kernel function.

Method scheme four: on the basis of method scheme three, the kernel function adopts a Gaussian kernel function.

Method scheme five: on the basis of method scheme one or two, the road centerline model constructed in said step 2) is:

E _p ＝∫{G[f(s)]} ² ds＝max

E _g =∫[f″(s)] ² ds=min

C _g ＝|f″(s)|≤T ₁

Among them, G(x) represents the road probability distribution function, f(s) represents the road centerline, and T ₁ is the set threshold.

Method 6: On the basis of method 5, the cost function constructed according to the road centerline model is:

Where p={p ₁ ,...,p _n }, p _i =( _xi ,y _i ) are the n vertices of the polyline segment on the center line of the road, S is the pixel set of the line segment p _i p _i+1 , is the gray function of the line segment p _i p _i+1 on the probability distribution graph, a _i is the direction of the line segment p _i p _i+1 , |Δs _i | is the length of the line segment p _i p _i+1 .

The present invention also provides a remote sensing image road extraction device, including the following six solutions, device solution 1: the road extraction device includes a memory, a processor, and a computer program stored in the memory and running on the processor , the processor is coupled to the memory, and the processor implements the following instructions when executing the computer program:

1) Calculate the road probability density of the original remote sensing image, and convert the road features on the original remote sensing image into the features of the road probability density;

2) Construct a road centerline model according to the characteristic that the probability value on the road centerline is higher than that on other positions, and determine the cost function according to the road centerline model;

3) Use dynamic programming to solve the maximum value of the cost function, and use this maximum value as the road centerline.

Device scheme two: on the basis of device scheme one, the road probability density of the remote sensing image in step 1) is determined by using support vector machine and kernel density estimation, and the specific process is as follows:

A. Using support vector machine to classify the remote sensing image, get a series of road sample points;

B. Calculate the probability density of each road sample point using kernel density estimation.

Device scheme three: on the basis of device scheme two, the probability density obtained in the step B is:

Where x _i is the i-th road sample point obtained, is the probability density estimate at point x, h is the bandwidth of the kernel density estimate, and K(x) is the kernel function.

Device scheme four: on the basis of device scheme three, the kernel function adopts a Gaussian kernel function.

Device scheme five: on the basis of device scheme one or two, the road centerline model constructed in said step 2) is:

E _p ＝∫{G[f(s)]} ² ds＝max

E _g =∫[f″(s)] ² ds=min

C _g ＝|f″(s)|≤T ₁

Among them, G(x) represents the road probability distribution function, f(s) represents the road centerline, and T ₁ is the set threshold.

Installation scheme six: On the basis of installation scheme five, the cost function constructed according to the road centerline model is:

Where p={p ₁ ,...,p _n }, p _i =( _xi ,y _i ) are the n vertices of the polyline segment on the center line of the road, S is the pixel set of the line segment p _i p _i+1 , is the gray function of the line segment p _i p _i+1 on the probability distribution graph, a _i is the direction of the line segment p _i p _i+1 , |Δs _i | is the length of the line segment p _i p _i+1 .

Description of drawings

Fig. 1 is the flow chart of the remote sensing image road extraction method of the present invention;

Fig. 2 is a flow chart of event processing of the remote sensing image road extraction device of the present invention.

detailed description

The specific embodiments of the present invention will be further described below in conjunction with the accompanying drawings.

The present invention divides remote sensing images into roads and non-roads by using support vector machine method according to the road pop features of high-resolution remote sensing images, obtains a series of road sample points, and uses kernel density estimation to calculate road probability distribution map; according to the road probability The probability value of the centerline on the distribution map is significantly higher than that of other positions to construct the road centerline model and cost function, and use dynamic programming to extract the road centerline on the remote sensing image. The process flow of this method is shown in Figure 1, and the specific implementation process is as follows:

1. Obtain remote sensing images, divide the remote sensing images into road and non-road, and obtain a series of road sample points.

Due to the problems of different objects in the same spectrum and same objects in different spectra in high-resolution remote sensing images, soft margin support vector machines were selected for classification. Support vector machine is a two-class classification model proposed by Corinna Cortes and Vapnik et al. in 1995. It shows many unique advantages in text classification, handwritten digit recognition, target recognition and face detection, and is considered to be the current One of the best learning algorithms. The present invention uses support vector machine to divide high resolution remote sensing image into road class and non-road class to obtain a series of road sample points. selected from the feature database.

2. Calculate the road probability distribution map of the obtained road sample points.

The road probability distribution map indicates the probability that each pixel on the remote sensing image is a road. The probability value is comprehensively calculated based on the pixel gray level and other factors. To deal with observation data that obeys an unknown distribution, it is usually necessary to estimate its probability from known data. density function, this is called a probability density estimate. At present, the commonly used probability density estimation methods include parameter estimation, histogram estimation and kernel density estimation. Although the probability density estimation based on the histogram can describe the inherent distribution of the data, there are still three main shortcomings: ① The graph is not smooth; ② The shape of the histogram is easily affected by the position of the starting point and the width of the interval; ③ When the data is Histogram estimation has significant limitations in three dimensions or higher. Kernel density estimation is a generalization of histogram estimation, but different from histogram estimation, kernel density estimation gives corresponding weights according to the degree of observation data close to the estimated point x, which overcomes the histogram estimation is not smooth, depends on the starting point, etc. shortcoming. The present invention uses kernel density estimation to calculate the probability distribution graph of each road sample point. The mathematical model used for kernel density estimation is:

Among them, x _i is the i-th road sample point, is the estimated value of the probability density at the road sample point x, h is the bandwidth of kernel density estimation, and K(x) is the kernel function. In kernel density estimation, commonly used kernel functions are Gaussian kernel function, Epanechnikov kernel function, triangular kernel function and rectangular kernel function. The present invention adopts Gauss kernel function, and its mathematical expression is similar to standard normal distribution:

3. Construct the road centerline model and cost function according to the characteristic that the probability value on the road centerline is significantly higher than that on other positions.

The present invention selects and utilizes the road feature on the road probability distribution map to construct the road centerline model and the cost function, and the road centerline model established according to the road probability distribution feature is:

1) On the road probability distribution map, the estimated probability value of the road sample point on the road centerline is larger than that of other road sample points, so the sum of the squares of the probability estimates of all sample points on the road centerline will reach A maximum value, namely:

E _p ＝∫{G[f(s)]} ² ds＝max

Among them, G(x) represents the road probability distribution function, and f(s) represents the road centerline.

2) According to the geometric characteristics of the road, the center line of the road should be a smooth curve, namely:

E _g =∫[f″(s)] ² ds=min

3) According to the requirements of traffic safety regulations, there is an upper bound on the local curve rate of the road, namely:

C _g ＝|f″(s)|≤T ₁

where T ₁ is a given threshold.

4. Use the dynamic programming algorithm to solve the road centerline model to realize the extraction of roads.

In the specific solution process, the road centerline is represented by a polyline segment containing n vertices, and the vertices on the polyline segment move around their initial positions ( _xi , y _i ), and the vertices of the polyline segment are p={p ₁ ,...,p _n },p _i ＝( _xi ,y _i ), the properties of the road centerline model 1) The discrete form is:

Where S is the pixel set of the line segment p _i p _i+1 , is the gray function of the line segment p _i p _i+1 on the probability distribution graph, and the discrete forms of properties 2) and 3) are:

C _g ＝|a _i -a _i+1 |<T ₁

Where a _i is the direction of line segment p _i p _i+1 , |Δs _i | is the length of line segment p _i p _i+1 , namely

According to the road centerline model, the following cost function is constructed:

Among them, the cost function E is the sum of a series of function items E _i , each E _i only depends on the three adjacent vertices {p _i-1 , p _i , p _i+1 } of the polyline segment, p _i =(x _i , y _i ), and the cost function E must satisfy the constraint condition |a _i -a _i+1 |<T ₁ .

Dynamic programming is mainly used to solve optimization problems, and the solution process using dynamic programming is:

For a cost function:

g=g(x ₁ ,x ₂ ,...,x _n ),0≤x _i ≤m _i ,i=1,2,...,n (1)

When its independent variables (x ₁ ,x ₂ ,...,x _n ) are discrete values and the cost function g is as follows:

g(x ₁ ,x ₂ ,...,x _n )＝g ₁ (x ₁ ,x ₂ ,x ₃ )+g ₂ (x ₂ ,x ₃ ,x ₄ )+...+g _n-2 (x _n-2 ,x _n-1 ,x _n ) (2)

The maximum value M of this function can be solved by dynamic programming algorithm, and the process is:

①For any variable x ₂ , x ₃ , solve the function f ₁ (x ₂ , x ₃ ):

②Continue to eliminate the variable x ₂ according to the first step, that is, for any variable x ₃ , x ₄ , solve the function f ₂ (x ₃ ,x ₄ ):

③Repeat the above steps to finally get:

Then the maximum value M of the cost function is:

Embodiment of a remote sensing image road extraction device of the present invention

The road extraction device in this embodiment includes a memory, a processor and a computer program stored on the memory and run on the processor, the processor is coupled with the memory, and the processor implements the following instructions when executing the computer program: 1) calculate the original remote sensing Image road probability density, transforming the road features on the original remote sensing image into the features of road probability density; 2) Constructing the road centerline model according to the feature that the probability value on the road centerline is higher than that on other positions, and according to the road The centerline model determines the cost function; 3) Use dynamic programming to solve the maximum value of the cost function, and use the maximum value as the road centerline. Processor can adopt single-chip microcomputer, DSP, PLC or MCU etc., memory can adopt RAM memory, flash memory, ROM memory, EPROM memory, EEPROM memory, register, hard disk, mobile disk, CD-ROM or any other form known in the art The storage medium and the specific implementation means of each instruction have been described in the embodiments of the method, and will not be repeated here. The event processing flow of the device is shown in FIG. 2 . The event flow during the implementation of the device is mainly divided into three parts, the main frame window, the main control entry of the remote sensing image road extraction device, and the database engine access component. First, the user performs corresponding operations through the main frame window to set the specific parameters of road collection; then the device invokes the data or models in the external database engine to obtain the corresponding road features; Road extraction is processed and the final result is evaluated.

The present invention can be applied to the road information collection operation of high-resolution remote sensing images, changing the existing geographic information guarantee mode, completing the road information collection tasks with a certain width on high-resolution remote sensing images through a small amount of manual participation, and significantly shortening the mapping cycle , greatly improving the automatic processing of remote sensing images. It plays an important role in my country's economic construction and social development. In addition, the present invention can also be used in inter-information monitoring of earthquake disasters, GIS updating, map drawing, etc., and has broad social application prospects and important application values.

Specific implementations have been given above, but the present invention is not limited to the described implementations. For those of ordinary skill in the art, according to the teachings of the present invention, it does not need to spend money to design various deformation models, formulas, and parameters. Creative work, changes, modifications, replacements and variations to the implementation without departing from the principle and spirit of the present invention still fall within the protection scope of the present invention.

Claims (10)

1. a kind of method for extracting remote sensing image road, it is characterised in that the method for extracting roads comprises the following steps：

1) original remote sensing image road probability density is calculated, it is close that the roadway characteristic on original remote sensing image is converted into road probability The feature of degree；

2) it is higher than the feature construction road-center line model of the probable value in other positions according to the probable value on road axis, And cost function is determined according to road-center line model；

3) maximum of cost function is solved using Dynamic Programming, road axis is used as using the maximum.

2. method for extracting remote sensing image road according to claim 1, it is characterised in that the remote sensing image of the step 1) Road probability density determines that detailed process is as follows using SVMs and Density Estimator：

A. remote sensing image is classified using SVMs, obtains series of road sample point；

B. the probability density of each road sample point is calculated using Density Estimator.

3. method for extracting remote sensing image road according to claim 2, it is characterised in that obtained in the step B general Rate density is：

<mrow> <mover> <mi>f</mi> <mo>^</mo> </mover> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <msup> <mi>nh</mi> <mi>d</mi> </msup> </mrow> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <mi>K</mi> <mrow> <mo>(</mo> <mfrac> <mrow> <mi>x</mi> <mo>-</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> </mrow> <mi>h</mi> </mfrac> <mo>)</mo> </mrow> </mrow>

Wherein x_iFor i-th obtained of road sample point,It is the probability density valuation at point x, h is the band of Density Estimator Width, K (x) are kernel function.

4. method for extracting remote sensing image road according to claim 1 or 2, it is characterised in that built in the step 2) Road-center line model be：

E_p=∫ { G [f (s)] }²Ds=max

E_g=∫ [f " (s)]²Ds=min

C_g=| f " (s) |≤T₁

Wherein G (x) represents road probability-distribution function, and f (s) represents road axis, T₁For given threshold.

5. method for extracting remote sensing image road according to claim 4, it is characterised in that according to road-center line model structure The cost function built is：

<mrow> <mi>E</mi> <mo>=</mo> <munder> <mo>&amp;Sigma;</mo> <mi>i</mi> </munder> <msub> <mi>E</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>p</mi> <mrow> <mi>i</mi> <mo>-</mo> <mn>1</mn> </mrow> </msub> <mo>,</mo> <msub> <mi>p</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>p</mi> <mrow> <mi>i</mi> <mo>+</mo> <mn>1</mn> </mrow> </msub> <mo>)</mo> </mrow> <mo>=</mo> <munder> <mo>&amp;Sigma;</mo> <mi>i</mi> </munder> <mo>&amp;lsqb;</mo> <mn>1</mn> <mo>+</mo> <mi>cos</mi> <mrow> <mo>(</mo> <msub> <mi>a</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mrow> <mi>i</mi> <mo>+</mo> <mn>1</mn> </mrow> </msub> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> <munder> <mo>&amp;Sigma;</mo> <mrow> <mi>s</mi> <mo>&amp;Element;</mo> <mi>S</mi> </mrow> </munder> <msub> <msup> <mi>g</mi> <mn>2</mn> </msup> <mrow> <msub> <mi>P</mi> <mi>i</mi> </msub> <msub> <mi>P</mi> <mrow> <mi>i</mi> <mo>+</mo> <mn>1</mn> </mrow> </msub> </mrow> </msub> <mrow> <mo>(</mo> <mi>s</mi> <mo>)</mo> </mrow> <mo>/</mo> <mo>|</mo> <msub> <mi>&amp;Delta;s</mi> <mi>i</mi> </msub> <mo>|</mo> </mrow>

Wherein p={ p₁,...,p_n},p_i=(x_i,y_i) be broken line on road axis n summit, S is line segment p_ip_i+1's Pixel set,For line segment p in probability distribution graph_ip_i+1Gamma function, a_iFor line segment p_ip_i+1Direction, | Δ s_i| it is line segment p_ip_i+1Length.

A kind of 6. remote sensing image road extraction element, it is characterised in that the road extraction device include memory and processor with And the computer program run on the memory and on the processor is stored in, the processor and the memory phase Coupling, realized described in the computing device during computer program to give an order：

1) original remote sensing image road probability density is calculated, it is close that the roadway characteristic on original remote sensing image is converted into road probability The feature of degree；

2) it is higher than the feature construction road-center line model of the probable value in other positions according to the probable value on road axis, And cost function is determined according to road-center line model；

3) maximum of cost function is solved using Dynamic Programming, road axis is used as using the maximum.

7. remote sensing image road extraction element according to claim 6, it is characterised in that the remote sensing image of the step 1) Road probability density determines that detailed process is as follows using SVMs and Density Estimator：

A. remote sensing image is classified using SVMs, obtains series of road sample point；

B. the probability density of each road sample point is calculated using Density Estimator.

8. remote sensing image road extraction element according to claim 7, it is characterised in that obtained in the step B general Rate density is：

<mrow> <mover> <mi>f</mi> <mo>^</mo> </mover> <mrow> <mo>(</mo> <mi>x</mi> <mo>)</mo> </mrow> <mo>=</mo> <mfrac> <mn>1</mn> <mrow> <msup> <mi>nh</mi> <mi>d</mi> </msup> </mrow> </mfrac> <munderover> <mo>&amp;Sigma;</mo> <mrow> <mi>i</mi> <mo>=</mo> <mn>1</mn> </mrow> <mi>n</mi> </munderover> <mi>K</mi> <mrow> <mo>(</mo> <mfrac> <mrow> <mi>x</mi> <mo>-</mo> <msub> <mi>x</mi> <mi>i</mi> </msub> </mrow> <mi>h</mi> </mfrac> <mo>)</mo> </mrow> </mrow>

Wherein x_iFor i-th obtained of road sample point,It is the probability density valuation at point x, h is the band of Density Estimator Width, K (x) are kernel function.

9. the remote sensing image road extraction element according to claim 6 or 7, it is characterised in that built in the step 2) Road-center line model be：

E_p=∫ { G [f (s)] }²Ds=max

E_g=∫ [f " (s)]²Ds=min

C_g=| f " (s) |≤T₁

Wherein G (x) represents road probability-distribution function, and f (s) represents road axis, T₁For given threshold.

10. remote sensing image road extraction element according to claim 9, it is characterised in that according to road-center line model The cost function of structure is：

<mrow> <mi>E</mi> <mo>=</mo> <munder> <mo>&amp;Sigma;</mo> <mi>i</mi> </munder> <msub> <mi>E</mi> <mi>i</mi> </msub> <mrow> <mo>(</mo> <msub> <mi>p</mi> <mrow> <mi>i</mi> <mo>-</mo> <mn>1</mn> </mrow> </msub> <mo>,</mo> <msub> <mi>p</mi> <mi>i</mi> </msub> <mo>,</mo> <msub> <mi>p</mi> <mrow> <mi>i</mi> <mo>+</mo> <mn>1</mn> </mrow> </msub> <mo>)</mo> </mrow> <mo>=</mo> <munder> <mo>&amp;Sigma;</mo> <mi>i</mi> </munder> <mo>&amp;lsqb;</mo> <mn>1</mn> <mo>+</mo> <mi>c</mi> <mi>o</mi> <mi>s</mi> <mrow> <mo>(</mo> <msub> <mi>a</mi> <mi>i</mi> </msub> <mo>-</mo> <msub> <mi>a</mi> <mrow> <mi>i</mi> <mo>+</mo> <mn>1</mn> </mrow> </msub> <mo>)</mo> </mrow> <mo>&amp;rsqb;</mo> <munder> <mo>&amp;Sigma;</mo> <mrow> <mi>s</mi> <mo>&amp;Element;</mo> <mi>S</mi> </mrow> </munder> <msub> <msup> <mi>G</mi> <mn>2</mn> </msup> <mrow> <msub> <mi>P</mi> <mi>i</mi> </msub> <msub> <mi>P</mi> <mrow> <mi>i</mi> <mo>+</mo> <mn>1</mn> </mrow> </msub> </mrow> </msub> <mrow> <mo>(</mo> <mi>s</mi> <mo>)</mo> </mrow> <mo>/</mo> <mo>|</mo> <msub> <mi>&amp;Delta;s</mi> <mi>i</mi> </msub> <mo>|</mo> </mrow>

Wherein p={ p₁,...,p_n},p_i=(x_i,y_i) be broken line on road axis n summit, S is line segment p_ip_i+1's Pixel set,For line segment p in probability distribution graph_ip_i+1Gamma function, a_iFor line segment p_ip_i+1Direction, | Δ s_i| it is line segment p_ip_i+1Length.