CN108491826B

CN108491826B - Automatic extraction method of remote sensing image building

Info

Publication number: CN108491826B
Application number: CN201810307327.1A
Authority: CN
Inventors: 施文灶; 程姗
Original assignee: Fujian Normal University
Current assignee: Fujian Normal University
Priority date: 2018-04-08
Filing date: 2018-04-08
Publication date: 2021-04-30
Anticipated expiration: 2038-04-08
Also published as: CN108491826A

Abstract

The invention relates to an automatic extraction method of a remote sensing image building. The method comprises the following steps: step 1, inputting a multispectral remote sensing image; step 2, traversing distances from all pixel points to two randomly selected points, and extracting two hierarchies; step 3, calculating the fluctuation intensity of the two layers; step 4, judging candidate building layers and non-building layers; step 5, initializing the number of the areas; step 6, processing the candidate building image layer; step 7, judging the iteration stop condition of the step 6; step 8, labeling processing; step 9, post-treatment; and step 10, outputting the result. The method can accurately extract buildings, especially dense buildings, in the multispectral remote sensing image, and can be applied to updating of the buildings in the urban geographical basic information database.

Description

Automatic extraction method of remote sensing image building

Technical Field

The invention relates to the field of remote sensing image processing, in particular to an automatic extraction method of a remote sensing image building.

Background

The building is one of main geographic elements of a city and is important content of various city thematic maps, and the research on the extraction of the building has important significance for comprehensively investigating the city geographic information environment. With the rapid development of the high-resolution remote sensing image acquisition technology, the remote sensing image has better data sources for processing, analyzing and applying, and the digital product has wider and deeper application. The computer image processing technology, the pattern recognition, the artificial intelligence and the like all make progress to different degrees, and the possibility is provided for efficiently extracting effective information in massive images. However, the building information is much more difficult to extract than other information such as roads and water bodies, and the main reasons are as follows:

(1) the data source is mainly a two-dimensional remote sensing image, and direct three-dimensional data is lacked in most cases;

(2) different remote sensing images often have larger difference due to different factors such as spectral range, resolution, geometric images of the sensor, imaging conditions and the like;

(3) the appearances, texture details and the like of different types of buildings are varied, the differences on remote sensing images are large, a unified building model base is difficult to establish, and automatic extraction of information is difficult;

(4) the complexity of the scene of the building, such as low contrast, mutual shielding of houses, shadows of the building itself, shadows of other objects, and the like, makes it difficult to automatically extract the building with clear boundaries from the background.

Disclosure of Invention

The invention provides an automatic extraction method of a remote sensing image building, which can overcome the problem of difficulty in extracting the building in the current remote sensing image, fully utilizes the characteristics of three components of R, G and B in the remote sensing image, can detect the building target in the remote sensing image based on the distance between characteristic vectors, does not need manual intervention and has high automation degree.

The technical scheme adopted for realizing the aim of the invention is as follows: the method comprises the following steps:

step 1: preprocessing an input multispectral remote sensing image containing three color components of R, G and B to obtain an image I_in(ii) a Step 2: in image I_inTwo points are randomly selected from the RGB color space SPA and are respectively marked as P₁And P₂Calculating an image I_inPixel point P in_xRespectively to P₁And P₂Are respectively marked as d_1xAnd d_2xWhen d is_1x≤d_2xThen, the pixel point P is set_xAnd P₁Are combined into the same set and marked as S₁Otherwise, the pixel point P is_xAnd P₂Are combined into the same set and marked as S₂When S is₁After adding new pixel, P is added₁Is updated to S₁Average position of all pixels (rounding the position average coordinate value) in S₂After adding new pixel, P is added₂Is updated to S₂Average position of all pixels in (rounding the position average coordinate value); iterating the above process of step 2 until the image I is traversed_inObtaining two layered layers of all the pixel points in the image₁And Layer₂；

And step 3: using the following formula to pair the two layered layers in step 2₁And Layer₂And (3) processing:

in the formula (1), Flt_{layer_num}Is a layered Layer_{layer_num}Intensity of fluctuation of (1), mean_{layer_num}Is a layered Layer_{layer_num}The mean value in the RGB color space SPA, layer _ num, is the two hierarchical numbers in step 2, the values are 1 and 2, I_iIs the value of the ith pixel in the RGB color space SPA, N_{layer_num}Is a layered Layer_{layer_num}The number of pixels in (1);

and 4, step 4: when Flt is present_{layer_num}Satisfies the condition T₀Then, the Layer will be layered_{layer_num}Judging as a candidate building Layer_cbOtherwise, Layer will be layered_{layer_num}Judging as non-building Layer_nb；

And 5: the Layer of the candidate building in the step 4 is Layer_cbIs initialized to C₀；

Step 6: using the following formula to perform Layer mapping on the candidate building_cbAnd (3) processing:

in the formula (2), ObjF is an objective function for region extraction, H_iIs the eigenvector, R, of the ith pixel point in the RGB color space SPA_iIs the value of the ith pixel point in the R component, G_iIs the value of the ith pixel point in the G component, B_iIs the value of the ith pixel point in the B component, H_QIs the characteristic vector of the centroid pixel point Qc of the region Q in the RGB color space SPA, R_QcIs the value of pixel point Qc in R component, G_QcIs the value of pixel point Qc in the G component, B_QcFor taking pixel point Qc from component BValue, p_Q(i) Is the probability that the ith pixel belongs to the region Q, | H_i-H_QI represents solving feature vector H_iAnd H_QThe distance between them;

and 7: when | ObjF^(k+1)-ObjF^(k)If | ≦ Thr, obtaining a region set RS, entering the step 8, otherwise, adding C₀Is updated to C₀+1, return to step 5, ObjF^(k)The value of objF after the kth iteration;

and 8: labeling the region set RS in the step 7;

and step 9: deletion area less than S₀The non-building area of (2) extracting the building with the squareness and the aspect ratio as constraints; step 10: and outputting and extracting a building result.

The preprocessing in step 1 includes geometric correction, radiation correction and contrast enhancement.

The distance H in the step 6_i-H_QThe method for obtaining the | | adopts the Chebyshev distance.

The invention has the beneficial effects that: the method can accurately extract buildings, especially dense buildings, in the multispectral remote sensing image, and can be applied to updating of the buildings in the urban geographical basic information database.

Drawings

FIG. 1 is an overall process flow diagram of the present invention.

Detailed Description

The following detailed description of embodiments of the invention refers to the accompanying drawings.

In step 101, a multispectral remote sensing image containing three color components of R, G and B is input.

In step 102, the input multispectral remote sensing image of step 101 is preprocessed, including geometric correction, radiation correction and contrast enhancement, to obtain an image I_in。

In step 103, in the image I_inTwo points are randomly selected from the RGB color space SPA and are respectively marked as P₁And P₂。

In step 104, image I is computed_inIn (1) imagePrime point P_xRespectively to P₁And P₂Are respectively marked as d_1xAnd d_2xWhen d is_1x≤d_2xThen, the pixel point P is set_xAnd P₁Are combined into the same set and marked as S₁Otherwise, the pixel point P is_xAnd P₂Are combined into the same set and marked as S₂When S is₁After adding new pixel, P is added₁Is updated to S₁Average position of all pixels (rounding the position average coordinate value) in S₂After adding new pixel, P is added₂Is updated to S₂The average position of all pixels in (rounding the position average coordinate value).

In step 105, it is determined whether to traverse the image I_inIf yes, two layered layers are obtained₁And Layer₂And proceeds to step 106, otherwise to step 104.

In step 106, the two layered layers in step 105 are processed using the following formula₁And Layer₂And (3) processing:

in formula (3), Flt_{layer_num}Is a layered Layer_{layer_num}Intensity of fluctuation of (1), mean_{layer_num}Is a layered Layer_{layer_num}The mean value in the RGB color space SPA, layer _ num, is the two hierarchical numbers in step 2, the values are 1 and 2, I_iIs the value of the ith pixel in the RGB color space SPA, N_{layer_num}Is a layered Layer_{layer_num}The number of pixels in (1).

In step 107, Flt is judged_{layer_num}Whether or not condition T is satisfied₀If yes, then Layer will be layered_{layer_num}Judging as a candidate building Layer_cbAnd proceeds to step 109, otherwise Layer will be layered_{layer_num}Judging as non-building Layer_nbAnd proceeds to step 108, where condition T is used to extract dense buildings based on experimental data₀Need to be set to Flt_{layer_num}Not less than 465, when used for extracting sparse buildings, condition T₀It is set to 250. ltoreq. Flt_{layer_num}＜465。

In step 108, a non-building Layer is obtained_nbThe input multispectral remote sensing image of step 101 is shown to contain no buildings.

In step 109, the candidate building Layer in step 107 is applied_cbIs initialized to C₀In order to take account of the running speed and the building extraction effect, C is added₀Set to 10.

At step 110, a region extraction objective function ObjF is constructed:

in the formula (4), ObjF is an objective function for region extraction, H_iIs the eigenvector, R, of the ith pixel point in the RGB color space SPA_iIs the value of the ith pixel point in the R component, G_iIs the value of the ith pixel point in the G component, B_iIs the value of the ith pixel point in the B component, H_QIs the characteristic vector of the centroid pixel point Qc of the region Q in the RGB color space SPA, R_QcIs the value of pixel point Qc in R component, G_QcIs the value of pixel point Qc in the G component, B_QcIs the value of the pixel point Qc in the B component, p_Q(i) Is the probability that the ith pixel belongs to the region Q, | H_i-H_QI represents solving feature vector H_iAnd H_QThe chebyshev distance therebetween.

In step 111, the area extraction objective function objF in step 110 is used to candidate building Layer_cbAnd (6) processing.

In step 112, using ObjF^(k)Representing the value of ObjF after the kth iteration, and judging whether the ObjF meets the stopping condition | ObjF^(k+1)-ObjF^(k)Thr is less than or equal to | if yes, a region set RS is obtained, the step 113 is entered, otherwise C is added₀Is updated to C₀+1 and go to step 111 for assuranceAccuracy of extraction result, threshold Thr is set to 10^-4。

In step 113, labeling processing is performed on the region set RS in step 112.

At step 114, post-processing is performed, including deleting areas less than S₀The non-building area of (2) is obtained by extracting a building using the squareness and the aspect ratio as constraints, and using S to prevent interference of a small area ground feature₀Set to 50.

In step 115, the result is output.

Claims

1. An automatic extraction method of a remote sensing image building is characterized by comprising the following steps:

step 1: preprocessing an input multispectral remote sensing image containing three color components of R, G and B to obtain an image I_in；

Step 2: in image I_inTwo points are randomly selected from the RGB color space SPA and are respectively marked as P₁And P₂Calculating an image I_inPixel point P in_xRespectively to P₁And P₂Are respectively marked as d_1xAnd d_2xWhen d is_1x≤d_2xThen, the pixel point P is set_xAnd P₁Are combined into the same set and marked as S₁Otherwise, the pixel point P is_xAnd P₂Are combined into the same set and marked as S₂When S is₁After new pixel is added, for S₁The average coordinate value of the positions of all the pixels is rounded and taken as P₁When new position of S₂After new pixel is added, for S₂The average coordinate value of the positions of all the pixels is rounded and taken as P₂The new location of (2); step 2 is executed iteratively until the image I is traversed_inObtaining two layered layers of all the pixel points in the image₁And Layer₂；

and 4, step 4: when Flt is present_{layer_num}If the interval falls within the interval, the condition is met;

in the formula (2), ObjF is an objective function for region extraction, H_iIs the eigenvector, R, of the ith pixel point in the RGB color space SPA_iIs the value of the ith pixel point in the R component, G_iIs the value of the ith pixel point in the G component, B_iIs the value of the ith pixel point in the B component, H_QIs the characteristic vector of the centroid pixel point Qc of the region Q in the RGB color space SPA, R_QcIs the value of pixel point Qc in R component, G_QcIs the value of pixel point Qc in the G component, B_QcIs the value of the pixel point Qc in the B component, p_Q(i) Is the probability that the ith pixel belongs to the region Q, | H_i-H_QI represents solving feature vector H_iAnd H_QThe distance between them;

and 7: when | ObjF^(k+1)-ObjF^(k)If | ≦ Thr, wherein Thr is an accuracy threshold of the extraction result, obtaining a region set RS, entering step 8, otherwise, adding C₀Is updated to C₀+1, return to step 5, ObjF^(k)The value of objF after the kth iteration;

and 8: labeling the region set RS in the step 7;

and step 9: deletion area less than S₀The non-building area of (2) extracting the building with the squareness and the aspect ratio as constraints;

step 10: and outputting and extracting a building result.

2. The method for automatically extracting the remote sensing image building according to claim 1, wherein the preprocessing in the step 1 comprises geometric correction, radiation correction and contrast enhancement.

3. The method for automatically extracting buildings according to claim 1, wherein the distance H in step 6_i-H_QThe method for obtaining the | | adopts the Chebyshev distance.