CN102103202A - Semi-supervised classification method for airborne laser radar data fusing images - Google Patents

Abstract

The invention relates to the technical field of airborne laser radar data processing, in particular to a semi-supervised classification method for airborne laser radar data fusing images. In the method, based on a semi-supervision concept, rough classification results of point cloud data are utilized to extract high-accuracy training sample data which is used for classifying high-resolution images; and in the post-processing process, based on target complexity, spurious building points are removed, the classification results are refined, and LiDAR (Laser Intensity Direction and Ranging) point cloud multiple characteristics are fused for cross validation so as to finally realize fine classification of airborne laser radar data. The method is a fusion classification method with high reliability and high classification accuracy. In the precondition without using near infrared data, the method achieves good effect of classification for point cloud tall vegetation and low vegetation areas.

Description

A kind of semi-supervised sorting technique of airborne laser radar data that merges image

Technical field

The present invention relates to the airborne laser radar data processing technology field, relate in particular to a kind of semi-supervised sorting technique of airborne laser radar data that merges image.

Background technology

Airborne LiDAR is a kind of novel active airborne remote sensing earth observation technology, can directly obtain the space three-dimensional point cloud information of target.Along with quickening of urbanization process, utilize airborne LiDAR technology to realize that the high-precision effect extraction of city terrestrial object information is significant, wherein basic and the most crucial technology is the classification of LiDAR cloud data.The point that is categorized as exposed ground is used for the generation of digital terrain model, for landform mapping, engineering survey, environmental planning etc. provide basic data; The point that is categorized as buildings and vegetation can be applicable to improve building model reconstruction in DTM model accuracy, the 3D digital city, urban green space research etc.Yet, because the cloud data that LiDAR provides can not directly obtain the semantic information (material and structure etc.) of body surface, be difficult to extract body information and topological relation, cause utilizing merely the airborne laser scan-data to carry out automatic intelligent intractabilities such as terrain classification and identification and strengthen.Existing airborne LiDAR point cloud Processing Algorithm is to the automatic decipher scarce capacity of complicated urban area landforms, is used for quality control and manual time in the practice to have occupied sizable ratio in whole data processing time.Therefore, need design automatic, efficient, healthy and strong LiDAR point cloud classification and modeling algorithm.

Existing studies show that, owing to lack corresponding texture and semantic information, the Classification and Identification of utilizing the airborne laser scan-data to carry out atural object separately has significant limitation with intelligent the processing, and is not high for the complex scene nicety of grading, can not satisfy actual classification and handle application demand.

Summary of the invention

Limitation at the single source remotely-sensed data classification of above-mentioned existence, the deficiency of especially single laser radar data nicety of grading, the purpose of this invention is to provide a kind of semi-supervised sorting technique of airborne laser radar data that merges image, utilize semi-supervised method training sample, utilize the classification of high resolution image and airborne LiDAR data fusion, finally reach the purpose of cloud data being carried out the high precision classification.

For achieving the above object, the present invention adopts following technical scheme:

Original laser radar data denoising step, this step adopt k nearest neighbor ball denoise algorithm, remove the noise that exists in the some cloud, and interpolation generate digital surface model DSM;

High accuracy number ground model DEM generates step, and the laser radar data of this step after to denoising adopt iteration triangulation network progressive encryption filtering method to obtain beating on the ground laser radar point, and interpolation generates high accuracy number ground model dem data;

The nDSM data generate step, and this step is subtracted each other original DSM data and dem data, obtains the nDSM data;

Laser radar data rough sort step, the non-ground point set that this step is obtained the filtering through the iteration triangulation network, at first by elevation information cut-point cloud, utilize local attribute to estimate again, constraints such as elevation obtain cloud data in the initial category information of 2 classifications such as high vegetation, buildings;

Based on the data-aided image classification training sample of LiDAR extraction step, the some cloud of this step after to rough sort at first carries out the elevation assignment according to classification information, and carries out the corresponding raster data of graticule mesh generation; Secondly manually choose seed points, by the seed mediated growth method acquisition sample areas of growing automatically; By the semi-automatic sample information of obtaining high vegetation, buildings, the exposed face of land of said method, as the high precision training sample of image classification;

It is that mask process is carried out in 0 zone that the classification step of associating nDSM mask, this step utilize the nDSM data to produce on the high resolution image behind the registration elevation;

Sorted pseudo-building object point is removed step, this step is because single classification results based on associating nDSM mask that relies on spectral information to carry out, can cause the mistake branch of buildings classification because of too much relying on spectral information, therefore utilize shape index and complexity to calculate and remove non-buildings categorical data, remove the laser spots of erroneous judgement for the building point;

Based on the step of the laser spots classification of image classification result and some cloud multiple characteristics cross validation, this step at first utilizes the classification step of associating nDSM mask and image classification result that sorted pseudo-building object point removal step process obtains that a cloud is carried out the classification assignment; Secondly utilize some cloud multiple characteristics (strength mean value, dispersion) and dem data etc. that the classification assigned result is verified again, the misclassification point of adjusting point cloud classification, cloud data is divided into the exposed face of land, 4 classifications such as low vegetation, high vegetation and buildings the most at last.

Described high accuracy number ground model DEM generates step and further comprises following substep:

1. raw data is carried out medium filtering and handle the utmost point low spot (noise spot that elevation is very low) of rejecting in the data;

2. the outsourcing rectangle of construction data, the height value on four summits of this outsourcing rectangle is set according to the arest neighbors criterion, then the outsourcing rectangle is carried out triangulation, and with it as initial landform surface model;

3. data are carried out the graticule mesh tissue, grid should be slightly larger than the size of maximum buildings, and wherein the minimum point in each grid is millet cake initially, and the millet cake of choosing is initially joined in the TIN (TIN);

4. calculate the angle on each leg-of-mutton distance of putting its place and it and an Atria summit, if the value that calculates then joins it in TIN less than the pre-set threshold condition;

5. repeat 4. up to there not being new point to join in the TIN;

6. interpolation generates DEM.

The classification step of described associating nDSM mask further comprises following substep:

1. based on the information of nDSM, to be divided into high regional peace face zone with a high resolution image of cloud registration, utilize described buildings and the high vegetation sample that gets access to based on the data-aided image classification training sample of LiDAR extraction step, elevation be 0 or the zone of elevation in assign thresholds adopt maximal possibility estimation to classify;

2. based on the information of nDSM, be 0 in the elevation zone or be lower than in the image capturing range of certain threshold value, manually choose the surface vegetation sample, utilize the described face of land high precision sample that obtains based on the data-aided image classification training sample of LiDAR extraction step simultaneously, adopt maximal possibility estimation to classify to obtain the exposed face of land and the short vegetation classification of spectral information of high precision classification for green.

The present invention has the following advantages and good effect:

1) to get access to the precision that is used for the image classification sample very high for the result of the present invention by the LiDAR rough sort.

2) the present invention is under the prerequisite of not using the near infrared data, reaches the good result that a cloud level vegetation and short vegetation area are classified.

Description of drawings

Fig. 1 is the synoptic diagram that concerns of noise spot and neighborhood on every side among the present invention.

Fig. 2 is an iteration triangulation network filtering synoptic diagram among the present invention.

Fig. 3 is the some cloud integrated classification process flow diagram based on semi-supervised classification and high resolution image provided by the invention.

Embodiment

A kind of semi-supervised sorting technique of airborne laser radar data that merges image provided by the invention, based on semi-supervised notion, utilize the rough sort result of cloud data, extract high precision training sample data, be used for the classification of high resolution image, and merge LiDAR point cloud multiple characteristics and carry out cross validation, realize that finally the essence of airborne laser radar data is classified.

The invention will be further described in conjunction with the accompanying drawings with specific embodiment below:

A kind of semi-supervised sorting technique of airborne laser radar data that merges image provided by the invention may further comprise the steps:

(1) original laser radar data denoising:

If the cloud data existence is starkly lower than or the utmost point low spot of projecting environment and aerial point, can considerable influence post-processing algorithm precision, therefore these noise spots of removal before data processing.This method is surveyed the noise of removing in the some cloud by setting up the k nearest neighbor ball, and at first data point set carries out space lattice and divides, and there is the space ball in imagination, and is the centre of sphere with current measuring point, and radius is got the distance of measuring point to 6 of place cube grids respectively.Get the space ball of radius minimum, in the grid that interferes with it, carry out the K-neighbor search, stop principle, then stop search if satisfy the search of being set up; Otherwise, thereby get the contiguous ball of K that the space ball of long radius is more set up point to be located.A cloud is being carried out in the process of noise processed, what mainly depend on point in the k nearest neighbor ball of point to be located and foundation judges apart from size whether this point to be located is noise.(as shown in Figure 1).

(2) high accuracy DEM generates

By utilizing the filtering of the iteration triangulation network to obtain the ground point set, ground point collection interpolation is obtained Grid DEM.Committed step is the filtering of the iteration triangulation network, the steps include: that 1. raw data being carried out medium filtering handles the utmost point low spot (noise spot that elevation is very low) (this step is finished in the 1st step) of rejecting in the data; 2. the outsourcing rectangle of construction data, the height value on four summits of this outsourcing rectangle is set according to the arest neighbors criterion, then the outsourcing rectangle is carried out triangulation, and with its as initial landform surface model (as among Fig. 2 a); 3. data are carried out the graticule mesh tissue, grid should be slightly larger than the size of maximum buildings, and wherein the minimum point in each grid is millet cake initially, and the millet cake of choosing is initially joined in the TIN (TIN); 4. calculate the angle on each leg-of-mutton distance of putting its place and it and an Atria summit, if the value that calculates then joins it (as b among Fig. 2) in TIN less than the pre-set threshold condition; 5. repeat 4. up to there not being new point to join (as the c among Fig. 2) in the TIN; 6. interpolation generates DEM.

Iteration triangulation network progressive encryption filtering algorithm principle as shown in Figure 2.

(3) nD SM (normalized digital surface model) data generate

Normalization digital surface model (nDSM) promptly carries out the data (being DSM-DEM) that obtain after the algebraic difference computing by digital surface model (DSM) and DEM, and nDSM can directly reflect the elevation information of atural object, has alleviated the elevation influence that topographic relief causes atural object.This processing is suitable for topographic relief usually and changes violent atural object areal coverage.

Generate graticule mesh DSM according to being inserted in the cloud data of above-mentioned principle with denoising, the ground point interpolation of data that utilizes the filtering of the iteration triangulation network to obtain generates Grid DEM, both data is subtracted each other again, thereby obtains the nDSM data.

(4) the automatic rough sort of cloud data

The non-ground point data that filtering is obtained are cut apart, the some cloud in the same plane is segmented in same section.Because buildings point obviously is higher than the point around it, and the intensity of variation of most of buildings roof surfaces is less, therefore to the laser spots after the Filtering Processing, by the comprehensive section of the cutting apart difference of elevation with ground around this section, and the localized variation degree on the described surface of the section of cutting apart (seeing that 4.1 and 4.2 save) is discerned this section of cutting apart and whether is belonged to construction zone.On the basis of removing construction zone, utilize the point set of elevation threshold value identification vegetation area.After automatic rough sort flow performing is finished, original LiDAR point cloud will be three other point sets of target class such as ground, buildings, vegetation by rough sort.

4.1. normal is estimated

Make the neighborhood of sample point p be (p1, p2 ..., pk),

Be the barycenter of the neighborhood of p, promptly

$\stackrel{\‾}{p}=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{p}_i$ Formula 1

Because each point in the some cloud all has x, y, three components of z, the covariance matrix of therefore putting p is one 3 * 3 a matrix, can be defined as

$C={\left[\begin{array}{c}{p}_1-\stackrel{\‾}{p}\\ \mathrm{\Λ}\\ p_k-\stackrel{\‾}{p}\end{array}\right]}^T\left[\begin{array}{c}{p}_1-\stackrel{\‾}{p}\\ \mathrm{\Λ}\\ p_k-\stackrel{\‾}{p}\end{array}\right]$ Formula 2

By the sample point in the p neighborhood that adds up to barycenter At the squared-distance of three component directions, covariance matrix C can describe the statistical property that these sample points distribute.

Consider the proper vector problem

Cv _j=λ _jV _jFormula 3

Because C is the positive semidefinite battle array of a symmetry, so all eigenwerts all should be real number values, proper vector v _jThen constitute vertical coordinate system, and corresponded respectively to three fundamental components of sample point set in the neighborhood.Eigenwert tolerance be sample point p in the neighborhood _i(i=1,2 ..., k) along the variation of individual features vector direction.

Suppose λ ₀≤ λ ₁≤ λ ₂, can draw to draw a conclusion: the plane

Be a such plane, it passes through center of mass point

And the abutment points that makes invocation point p arrives the squared-distance and the minimum on this plane.Can think that also plane T (x) is curved surface approaching in a section at p place.Therefore, vector v ₀What can be similar to regards the surface normal n that approaches a p place as _p, vector v 1 and v2 have then generated curved surface in a section at p place.

4.2 Curvature Estimation

Based on 4.1 normal methods of estimation, utilize in the neighborhood normal of sample point to estimate this curvature on curved surface.Suppose λ ₀≤ λ ₁≤ λ ₂, tolerance be of the variation of the neighborhood of a p along the surface normal direction, promptly abutment points departs from the degree of section Tp.The overall departure degree of abutment points, i.e. abutment points p _iWith the square distance of barycenter with can provide by following formula:

$\underset{i=1}{\overset{k}{\mathrm{\Σ}}}{|p_i-\stackrel{\‾}{p}|}^2={\mathrm{\λ}}_0+{\mathrm{\λ}}_1+{\mathrm{\λ}}_2$ Formula 4

Therefore, be under the condition of k in the neighborhood size, the curved surface at some p place changes and can be defined as

${\mathrm{\σ}}_k\left(p\right)=\frac{{\mathrm{\λ}}_0}{{\mathrm{\λ}}_0+{\mathrm{\λ}}_1+{\mathrm{\λ}}_2}$ Formula 5

If σ _k(p)=0, show that then all points are all on the Tp of section.When these under the variation on all directions all is identical situation, curved surface changes and reaches its maximal value 1/3.Curved surface changes and can change to some extent along with the difference of selected neighborhood size.When the neighborhood value bigger the time, estimated curved surface changes just violent, when the neighborhood value was smaller, curved surface changed just more smooth.

(5) extract based on the data-aided image classification training sample of LiDAR

Some cloud after the rough sort, carry out the corresponding raster data of graticule mesh generation after the elevation assignment respectively according to classification information. show this raster data by elevation, and artificial interpretation selects in certain category regions in the grid image point as seed points, carry out the seed points growth method, arrive any pixels in the zone by upper and lower, left and right, upper left, lower-left, upper right and eight directions in bottom right, thereby obtain this sample areas.By these means, the ground classification, the buildings classification, the sample information of high vegetation classification will accurately be gathered.

(6) classification of associating nDSM mask.

A: based on the information of nDSM, to be divided into high regional peace face zone with a high resolution image of cloud registration, the buildings and the high vegetation sample that utilize (5) step to get access to, elevation greater than 0 or elevation adopt maximal possibility estimation to classify greater than the imagery zone in the assign thresholds.When using image data merely usually, artificial ground and buildings, meadow and high vegetation are caused interference easily mutually, cause the wrong branch of classification, but carry out mask owing to introduce the nDSM elevation information, separated the classification of disturbing each other, so the precision of the sorting result of buildings and high vegetation improves.

: based on the information of nDSM, be 0 in the elevation zone or be lower than in the image capturing range of certain threshold value, manually choose the surface vegetation sample, the face of land high precision sample that utilizes (5) step to obtain simultaneously, the exposed face of land and the short vegetation classification of spectral information of adopting maximal possibility estimation to classify to obtain the high precision classification for green.

(7) sorted pseudo-building object point is removed

Because the classification of single dependence spectral information causes buildings classification mistake to be divided easily, therefore need revise these mistakes and divide.

Buildings can be interpreted as geometric properties target unified, that have certain meaning usually.Therefore can remove as the non-building object point that shape index (as formula 6) comes mistake to be divided into the buildings class according to the geometric properties of target.The size of building target is represented by area or girth usually.Its region area can be represented with the number of pixels of compositing area; Its girth obtains by the length of computation bound line, i.e. the computation bound length of curve.The definition of shape index such as formula 6, wherein S is a target area, P is a girth.The common shape index of non-building target is less, and the building target shape is complicated more, and shape index numerical value is big more.

$I=\frac{\sqrt{S}}{P}$ Formula 6

(8) classify based on the laser spots of image classification result and some cloud multiple characteristics cross validation

Be used for the classification information of cloud data after the rough sort is carried out assignment again by the atural object classification figure that obtains after (6) and (7) step process, but the classification information that relying on classification figure merely provides can cause some zone of buildings and high vegetation to produce wrong the branch, therefore utilizing classification figure that a cloud is carried out a minute time-like, need utilize the various features of cloud data to come the classification information of a cloud is verified again, promptly by taking all factors into consideration the category attribute that features such as elevation, intensity, dispersion decide final laser spots.

A, utilize the wrong branch among the elevation character recognition atural object classification figure

Therefore the elevation of buildings and high vegetation point should can be used as a standard and distinguish the ground point that mistake is divided into buildings or high vegetation classification far above its peripheral ground point in theory.Therefore when a point is judged as buildings or high vegetation classification according to the classification information of classification figure, should utilize the coordinate figure of this point to carry out interpolation in dem data, whether the difference of the elevation that obtains according to this point relatively and the elevation of this point itself judges that greater than pre-set threshold whether this put ground point.If greater than assign thresholds, this meets this decision theory away from ground so, can keep original classification information constant.Otherwise the classification information of this point then should be ground point, and can directly give this classification is the open ground object point.

B, utilize strength characteristic to distinguish wrong branch among the atural object classification figure

Buildings strength information and high vegetation cover strength are completely different, with the standard of intensity as a differentiation, to the high vegetation point classification information correction of excepting for the buildings class.Some cloud intensity level in to classification figure in construction zone and the high vegetation area is added up mean value BI and the TI that obtains separately respectively, marginal point to construction zone travels through, if (threshold value is BI ± BI/4) to the intensity of a buildings classification point usually not in the buildings threshold range of setting, and be in the high vegetation cover strength threshold range (threshold value is TI usually) simultaneously, then this point has and greatly may belong to high vegetation point, this laser spots should be divided into high vegetation point, otherwise keep former classification information constant.

C, utilize the wrong branch among the dispersion character recognition atural object classification figure

The spatial spreading degree of laser point cloud is an important clue of distinguishing buildings and vegetation.Because buildings and exposed ground are made up of the plane usually, it can be thought in the dispersion degree in space and distributes along two-dimensional surface (not necessarily level), and the point on the trees all comparatively disperses on all directions of space, and this discreteness can be analyzed by the eigenwert of discrete matrix.

Concrete grammar is as follows: the whole consecutive point in the search laser spots neighborhood, set up this laser spots 3 * 3 discrete matrix spatially, and see formula 7:

$S_j=\underset{i=0}{\overset{n}{\mathrm{\Σ}}}\left(v_i^T{v}_i\right)(j=\mathrm{0,1,2}.....M)$ Formula 7

S wherein _jBe 3 * 3 discrete matrix of j point, n is a consecutive point number in j the vertex neighborhood, v _iBe the volume coordinate v of the i consecutive point of j point _i=(x _i, y _i, z _i),

Be v _iTransposed matrix, M is that laser is counted.

With discrete matrix S _jMake svd, can obtain three eigenwerts of this dot matrix, and eigenwert is arranged from small to large.Set three classifications:

A, individual eigenwert are much larger than the another one eigenwert, and then this point is marked as the plane class

B, an eigenwert are much larger than two other eigenwert, and then this point is marked as the edge class

C, three eigenwerts then are labeled as the spatial spreading class all enough greatly.

Based on above-mentioned three standards, traversal vegetation point also utilizes dispersion to correct the wrong building object point that is divided into vegetation.

Above embodiment is only for the usefulness that the present invention is described, but not limitation of the present invention, person skilled in the relevant technique; under the situation that does not break away from the spirit and scope of the present invention; can also make various conversion or modification, so all technical schemes that are equal to, all fall into protection scope of the present invention.

Claims (3)

1. the semi-supervised sorting technique of airborne laser radar data that merges image is characterized in that, may further comprise the steps:

Original laser radar data denoising step, this step adopt k nearest neighbor ball denoise algorithm, remove the noise that exists in the some cloud, and interpolation generate digital surface model DSM;

High accuracy number ground model DEM generates step, and the laser radar data of this step after to denoising adopt iteration triangulation network progressive encryption filtering method to obtain beating on the ground laser radar point, and interpolation generates high accuracy number ground model dem data;

The nDSM data generate step, and this step is subtracted each other original DSM data and dem data, obtains the nDSM data;

Laser radar data rough sort step, the non-ground point set that this step is obtained the filtering through the iteration triangulation network, at first by elevation information cut-point cloud, utilize local attribute to estimate again, constraints such as elevation obtain cloud data in the initial category information of 2 classifications such as high vegetation, buildings;

Based on the data-aided image classification training sample of LiDAR extraction step, the some cloud of this step after to rough sort at first carries out the elevation assignment according to classification information, and carries out the corresponding raster data of graticule mesh generation; Secondly manually choose seed points, by the seed mediated growth method acquisition sample areas of growing automatically; By the semi-automatic sample information of obtaining high vegetation, buildings, the exposed face of land of said method, as the high precision training sample of image classification;

The classification step of associating nDSM mask, this step are utilized the nDSM data to produce on the high resolution image behind the registration elevation is lower than the plane domain of specifying the elevation threshold value and elevation and are higher than the high zone of specifying the elevation threshold value and carry out the mask process utilization,

Sorted pseudo-building object point is removed step, this step be because single rely on that spectral information carries out based on associating nDSM mask classification results, can be because of the simple mistake branch that relies on spectral information to cause the buildings classification, therefore utilize shape index and complexity to calculate and remove non-buildings categorical data, remove the laser spots of erroneous judgement for the building point;

Based on the step of the laser spots classification of image classification result and some cloud multiple characteristics cross validation, this step at first utilizes the classification step of associating nDSM mask and image classification result that sorted pseudo-building object point removal step process obtains that a cloud is carried out the classification assignment; Secondly utilize some cloud multiple characteristics (strength mean value, dispersion) and dem data etc. that the classification assigned result is verified again, the misclassification point of adjusting point cloud classification, cloud data is divided into the exposed face of land, 4 classifications such as low vegetation, high vegetation and buildings the most at last.

2. the semi-supervised sorting technique of the airborne laser radar data of fusion image according to claim 1 is characterized in that:

Described high accuracy number ground model DEM generates step and further comprises following substep:

1. raw data is carried out medium filtering and handle the utmost point low spot (noise spot that elevation is very low) of rejecting in the data;

2. the outsourcing rectangle of construction data, the height value on four summits of this outsourcing rectangle is set according to the arest neighbors criterion, then the outsourcing rectangle is carried out triangulation, and with it as initial landform surface model;

3. data are carried out the graticule mesh tissue, grid should be slightly larger than the size of maximum buildings, and wherein the minimum point in each grid is millet cake initially, and the millet cake of choosing is initially joined in the TIN (TIN);

4. calculate the angle on each leg-of-mutton distance of putting its place and it and an Atria summit, if the value that calculates then joins it in TIN less than the pre-set threshold condition;

5. repeat 4. up to there not being new point to join in the TIN;

6. interpolation generates DEM.

3. the semi-supervised sorting technique of the airborne laser radar data of fusion image according to claim 1 and 2 is characterized in that:

The classification step of described associating nDSM mask further comprises following substep:

1. based on the information of nDSM, to be divided into high regional peace face zone with a high resolution image of cloud registration, utilize described buildings and the high vegetation sample that gets access to based on the data-aided image classification training sample of LiDAR extraction step, elevation greater than 0 or elevation adopt maximal possibility estimation to classify greater than the zone of assign thresholds;

2. based on the information of nDSM, be 0 in the elevation zone or be lower than in the image capturing range of certain threshold value, manually choose the surface vegetation sample, utilize the described face of land high precision sample that obtains based on the data-aided image classification training sample of LiDAR extraction step simultaneously, adopt maximal possibility estimation to classify to obtain the exposed face of land and the short vegetation classification of spectral information of high precision classification for green.

Images