CN111428784A - Robust segmentation method for deciduous forest tree-level parameter measurement by adopting airborne laser radar - Google Patents
Robust segmentation method for deciduous forest tree-level parameter measurement by adopting airborne laser radar Download PDFInfo
- Publication number
- CN111428784A CN111428784A CN202010206026.7A CN202010206026A CN111428784A CN 111428784 A CN111428784 A CN 111428784A CN 202010206026 A CN202010206026 A CN 202010206026A CN 111428784 A CN111428784 A CN 111428784A
- Authority
- CN
- China
- Prior art keywords
- trunk
- tree
- points
- crown
- parameters
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 127
- 230000011218 segmentation Effects 0.000 title claims abstract description 61
- 238000005259 measurement Methods 0.000 title claims description 11
- 238000007781 pre-processing Methods 0.000 claims abstract description 19
- 239000010410 layer Substances 0.000 claims description 32
- 238000001514 detection method Methods 0.000 claims description 29
- 238000009499 grossing Methods 0.000 claims description 7
- 238000000605 extraction Methods 0.000 claims description 6
- 238000012805 post-processing Methods 0.000 claims description 3
- 239000002344 surface layer Substances 0.000 claims description 3
- 239000007943 implant Substances 0.000 claims description 2
- 238000012544 monitoring process Methods 0.000 abstract description 3
- 238000009826 distribution Methods 0.000 description 5
- 210000000481 breast Anatomy 0.000 description 4
- OKTJSMMVPCPJKN-UHFFFAOYSA-N Carbon Chemical compound [C] OKTJSMMVPCPJKN-UHFFFAOYSA-N 0.000 description 3
- 240000003793 Rhizophora mangle Species 0.000 description 3
- 229910052799 carbon Inorganic materials 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 3
- 244000074209 Leptadenia hastata Species 0.000 description 2
- 235000001008 Leptadenia hastata Nutrition 0.000 description 2
- 241000219492 Quercus Species 0.000 description 2
- 244000274906 Quercus alba Species 0.000 description 2
- 235000009137 Quercus alba Nutrition 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 2
- 238000011161 development Methods 0.000 description 2
- 230000018109 developmental process Effects 0.000 description 2
- 241000894007 species Species 0.000 description 2
- 241000208140 Acer Species 0.000 description 1
- 238000012935 Averaging Methods 0.000 description 1
- 239000002028 Biomass Substances 0.000 description 1
- 240000004385 Centaurea cyanus Species 0.000 description 1
- 235000005940 Centaurea cyanus Nutrition 0.000 description 1
- 235000008418 Hedeoma Nutrition 0.000 description 1
- 241000218314 Liriodendron tulipifera Species 0.000 description 1
- 244000246386 Mentha pulegium Species 0.000 description 1
- 235000016257 Mentha pulegium Nutrition 0.000 description 1
- 240000004885 Quercus rubra Species 0.000 description 1
- 235000017368 Quercus rubra var. ambigua Nutrition 0.000 description 1
- 235000017499 Quercus rubra var. rubra Nutrition 0.000 description 1
- 241001092459 Rubus Species 0.000 description 1
- 240000003021 Tsuga heterophylla Species 0.000 description 1
- 238000005452 bending Methods 0.000 description 1
- 239000005312 bioglass Substances 0.000 description 1
- 238000007621 cluster analysis Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000002349 favourable effect Effects 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 238000012417 linear regression Methods 0.000 description 1
- 239000003550 marker Substances 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000005192 partition Methods 0.000 description 1
- 238000012545 processing Methods 0.000 description 1
- 238000000611 regression analysis Methods 0.000 description 1
- 238000000926 separation method Methods 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000009827 uniform distribution Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- 239000002023 wood Substances 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/23—Clustering techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/22—Matching criteria, e.g. proximity measures
-
- G06T5/70—
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/11—Region-based segmentation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06T—IMAGE DATA PROCESSING OR GENERATION, IN GENERAL
- G06T7/00—Image analysis
- G06T7/10—Segmentation; Edge detection
- G06T7/136—Segmentation; Edge detection involving thresholding
Abstract
A robust segmentation method for determining deciduous forest tree-level parameters by using an airborne laser radar relates to the technical field of forest resource monitoring and comprises the following steps of scanning a low-altitude flying laser radar and a hollow flying laser radar in a deciduous forest region, respectively acquiring L eaf-on data and L eaf-off data, preprocessing and analyzing the L eaf-on data, then obtaining tree crown parameters by using a multi-threshold segmentation method, preprocessing and analyzing the L eaf-off data, separating all trunk radar points, clustering the obtained trunk radar points to obtain trunk parameters, matching the tree crown parameters with the trunk parameters to obtain single trees, finally segmenting and describing the single trees, and extracting tree-level structure parameters from the single trees to complete determination of the corresponding deciduous forest tree-level parameters.
Description
Technical Field
The invention relates to the technical field of forest resource monitoring, in particular to a robust segmentation method for determining deciduous forest tree-level parameters by adopting an airborne laser radar.
Background
Traditionally, forest surveys are labor intensive, time consuming, and limited by space accessibility, as most tree-level information and parameters are estimated by manually measuring sample plots in field surveys.
The method for obtaining the three-dimensional (3D) vegetation structure through the airborne laser radar has excellent capacity, and particularly, after single trees are extracted from remote sensing data (namely point cloud generated by the laser radar), the height of the trees and the circumference or diameter of the tree crowns can be directly extracted, the positions of the single trees are estimated according to the peak positions of the tree crowns, and then the factors such as the diameter at breast height, volume of timber, biomass, carbon storage and the like can be further calculated by combining with a different-speed growth equation or regression analysis.
In order to extract tree structure information from airborne lidar data, various methods or algorithms have been proposed. Early methods utilized pre-processed data of a Digital Surface Model (DSM) or a Canopy Height Model (CHM) to identify singles. The DSM-based method locates the global maximum elevation between the lidar surface points, generates a vertical profile, and creates a convex hull to delineate the crown. The CHM is derived from the differences between canopy surface height and Digital Elevation Model (DEM), and CHM-based methods detect treetops using local maximum filters and then delineate individual crowns using marker-controlled watershed segmentation schemes, region growing techniques, or flooding techniques. Since the CHM-based algorithm is essentially a raster image interpolated from spatially discrete points describing the vegetation canopy surface, there may be inherent errors and uncertainties therein. Spatial errors introduced during interpolation can reduce the accuracy of tree segmentation and tree-level information extraction. Generally, these methods have an inherent disadvantage in that only surface lidar data is considered when identifying and delineating singles, thereby missing under-forest trees.
To overcome the disadvantages of the above algorithm, the present applicant disclosed in chinese patent application No. 2019103511285, filed on 28.4.2019, a "method for determining tree-level parameters of deciduous forest based on small spot airborne radar", by a multi-threshold segmentation method, tree segmentation is performed on the upper and lower forests using different thresholds, and then the segmented tree segments are combined across canopy layers, which method results in an improvement in the average detection rate and average total accuracy of the under forest vegetation and the total accuracy of all trees, but in practical applications, it is not easy to identify and delineate under forest vegetation having a complex structure, which is due to overlapping of point cloud patches caused by complex terrain and vegetation conditions, when crown segmentation is performed on L iDAR point clouds, a cluster of trees that are very close to each other may be merged into one, but they are identified as multi-leaf trees in field segmentation.
Disclosure of Invention
The invention aims to provide a robust segmentation method for deciduous forest tree-level parameter determination by adopting an airborne laser radar, which improves the accuracy of tree-level parameter extraction by combining multi-threshold segmentation and a trunk detection method.
In order to solve the technical problem, the robust segmentation method for determining the deciduous forest tree-level parameters by using the airborne laser radar comprises the following steps of scanning a low-altitude flying laser radar and a hollow flying laser radar in a deciduous forest area, acquiring L eaf-on data and L eaf-off data respectively, preprocessing and analyzing the L eaf-on data, acquiring tree crown parameters by using a multi-threshold segmentation method, preprocessing and analyzing the L eaf-off data, separating all trunk radar points, clustering the obtained trunk radar points to obtain trunk parameters, matching the tree crown parameters with the trunk parameters to obtain single trees, segmenting and depicting the single trees, and extracting tree-level structure parameters from the single trees to complete determination of the corresponding deciduous forest tree-level parameters.
Preferably, the method of pre-processing the L eaf-on data or L eaf-off data comprises the steps of:
(1) dividing laser radar points in corresponding data into ground points and non-ground points, generating a Digital Elevation Model (DEM) with the resolution of 1 meter based on the ground points, performing gap filling by adopting a nearest neighbor method, and performing interpolation by adopting an average method;
(2) normalizing radar spot spacing and creating a resolution grid equal to a nominal spot spacing (NPS);
(3) selecting each grid cell as a highest elevation point of a lidar surface point (L SP) to filter the lidar point cloud;
(4) calculating the ground level of all L SPs using a lidar derived DEM;
(5) the L SP was smoothed using a gaussian smoothing filter with a standard deviation equal to NPS and a radius of 3 × NPS to reduce small variations in canopy implant height.
More preferably, the method for performing multi-threshold segmentation after the L eaf-on data are preprocessed comprises the following steps:
(1) layering the point cloud according to the echo number of the laser radar;
(2) performing a multi-threshold segmentation within each separated canopy;
(3) combining crown sections in a cross-layer manner;
(4) the resulting individual crowns of the segmentation are depicted.
More preferably, the method for separating all trunk radar points after preprocessing the L eaf-off data comprises the following steps:
(1) splitting a lidar point cloud in a sample plot into N with a height of 30 meterslA layer;
(2) calculating laser radar point n of each layeriThe number of laser radar points and the percentage of the total number of laser radar points in each layer in the sample plot;
(3) forming a histogram in which points are distributed in the height direction;
(4) the density of search points exceeds a predetermined threshold hbaseThe lowest level layer l ofx；
(5) Will lxIs defined as a dividing plane,/xThe following points are then potential trunk radar points, including one or more trunks.
More preferably, the method for clustering the trunk radar points to obtain the trunk parameters is to cluster the trunk radar points according to the spatial neighborhood relationship thereof to obtain the estimated trunk number and position in one sample plot, and allocate the trunk radar points to the trunk point cluster.
More preferably, the DBSCAN-based trunk detection and identification method is used to eliminate points on shrubs or isolated branches in the trunk radar points and to draw trunks of different shapes, wherein the DBSCAN-based trunk detection and identification method comprises the following steps:
(1) assigning the trunk radar points to estimated trunks without presupposing trunk shapes or numbers, wherein the trunk points are all projected onto a horizontal plane;
(2) the clustering process based on the DBSCAN is realized in a two-dimensional space formed by projection points on a horizontal plane of the previous step by using a group of parameters (Eps, MinPts), wherein Eps represents a clustering radius, MinPts represents the number of points on a circle, namely a clustering density threshold, the initial values of the two parameters are calculated by the formula (1) and the formula (2), and then the optimal value in a sample plot is obtained through multiple iterative comparisons;
more preferably, the trunk point cluster is subjected to data post-processing, a point cluster with the minimum height greater than 4m, a point cluster with the maximum height less than 2m and a trunk point cluster with the quarter-bit distance less than 1m are excluded, and trunk parameters are finally obtained.
More preferably, the method for matching the crown parameters and the trunk parameters to obtain the single wood comprises the following steps:
(1) given forest canopy number ilay( i lay1, 2, 3), the detected trunk number is jstem(jstem＝0，1，2，…)；
(2) Projecting the segmented crown point clusters in each layer onto an x-y plane, and checking a trunk extraction result related to the crown point segments;
(3) according to the detected number of the trunks, the matching method of the crown sections and the trunks comprises the following conditions:
(3.1) when the projected point of the trunk is not contained in the crown part in a certain projection plane, namely the crown part is regarded as a branch and is merged to the tree closest to the branch;
(3.2) when a trunk appears in a certain projection layer to be contained in a crown section, enabling the trunk and the crown section in the given crown layer to be used as a candidate matching pair, and further determining whether the point clusters are combined together to outline the whole structure of the single tree;
(3.3) when two or more trunks are contained in one crown section in a certain projection surface layer, judging that the plurality of small trees are completely in the vertical projection range of the crown of one large tree;
(3.4) when a case where no crown section is present in the region directly above the trunk of a certain projected course, it is judged that no crown section belongs to it, and it is regarded as a dry tree or a low shrub and deleted from the trunk point cluster.
More preferably, the method for segmenting and describing the single tree and extracting tree-level structure parameters from the single tree is that after the segmented single tree is determined through matching of the crown section point cluster and the trunk point cluster, the tree height, the crown size, the trunk position and the number of trees in the sample plot in the tree structure parameters can be directly derived from the segmented point clusters.
When the structural parameters of the tree are derived, the highest point in the point cluster of each laser radar derived tree is regarded as the top point of the tree, the vertical distance between the highest point and the ground is regarded as the height of the tree, and the perimeter of the tree crown is defined as the length of a boundary line of the tree crown projected onto an x-y plane.
The invention has the beneficial effects that: the method provided by the invention can detect more trees in a sample plot compared with the conventional multi-threshold method and DSM-based method, and the average detection rate, the accuracy rate of detecting trees and the total segmentation accuracy under the condition of considering CE (entrusting error) and OE (missing error) are obviously improved compared with the average detection rate and the accuracy rate of detecting trees and the total segmentation accuracy under the condition of considering CE (entrusting error) and OE (missing error), and the data are also obviously improved in the statistical results of the segmentation number of the upper-layer trees and the lower-layer trees.
Therefore, the trunk detection result in the leaf-off data is combined with the crown point segmented in the leaf-on data, so that the trunk detection result and the crown point segmented in the leaf-on data can be mutually referred, omission and misoperation are reduced, the whole tree detection precision is improved, and the mode of combining the data is obviously superior to the mode of using one data to perform tree-level segmentation. Furthermore, this also allows a significant improvement in the position of the correctly detected tree, since the determined trunk cluster points are more accurate than the vertices derived from the laser radar surface maxima.
The method provided by the invention is actually used as a method for tree trunk detection-assisted multi-threshold segmentation, has robustness on single tree segmentation and tree-level parameter extraction, is particularly suitable for deciduous forests, and is beneficial to the rapid development of the single tree-based method in the aspect of airborne laser radar data application so as to improve the accuracy and efficiency of forest resource clearing.
Drawings
FIG. 1 is a processing framework for application of the method provided in an embodiment of the invention;
FIG. 2 is four routines in the multi-threshold tree splitting method of an embodiment;
FIG. 3 is a flow chart of DBSCAN clustering algorithm in the embodiment;
FIG. 4-a is a diagram of a crown point cloud within a sample plot to isolate the distribution of leaf-on and leaf-off radar points among potential trunk points;
FIG. 4-b illustrates a use of a leaf-off point cloud to determine a separation plane (where N is the number of possible stem points in a sample plot) in separating potential crown point clouds i10 horizontal layers, hbase＝5％)；
FIG. 5-a is a side view of potential trunk points isolated in a plot after trunk detection and identification using a clustering method based on DBSCAN;
FIG. 5-b is a three-dimensional view of potential stem points isolated in a plot after stem detection and identification using a clustering method based on DBSCAN;
FIG. 5-c is a side view of a trunk cluster point cluster detected in a plot after trunk detection and identification using a DBSCAN-based clustering method;
FIG. 5-d is a three-dimensional view of a cluster of trunk cluster points detected in a sample plot after trunk detection and identification using a clustering method based on DBSCAN;
FIG. 6-a is a combined illustration of matching trunk and crown sections and vertically adjacent tier sections, where white dots represent the vertices of each crown section and crosses represent projected points of the trunk detected within each tier;
FIG. 6-b is a representation of a junction growth tree (CGT) in which the root nodes of each tree are populated with grids, such as R5 and R8;
FIG. 7 is a flow chart of matching crown segments to trunks in an embodiment;
FIG. 8-a is a result of using only leaf-on point cloud data in a prior art multi-threshold segmentation method;
FIG. 8-b is a cross-layer combination rule and lidar derived tree when using existing multi-threshold segmentation methods;
8-c are matching rules for trunk and crown segments under the method provided in an embodiment of the invention;
FIG. 8-d is a segmentation result of point cloud data using lead-on and leaf off simultaneously according to the method provided in the embodiment of the present invention;
FIG. 9 is a ratio of estimated tree height to reference tree height for all correctly detected trees using method (a) provided by the present invention, using multi-threshold segmentation method (b), using DSM-based method (c) (where the dashed line represents the assumed 1:1 ratio between the radar-derived tree and the reference tree);
fig. 10 shows the distance distribution between the radar-derived tree positions and the reference tree positions of all correct detection trees obtained by the method (a) provided by the present invention, using two data (trunk and vertex) as the trunk map, the method (b) of multi-threshold segmentation, and the method (c) based on DSM.
Detailed Description
In order to facilitate understanding of those skilled in the art, the present invention will be further described with reference to the following examples and drawings, which are not intended to limit the present invention.
The site of implementation of this example is robinson forest (RF, coordinates: 37 ° 28 '23 "N83 ° 08' 36" W) at the university of kentucky, eastern part of britain, rugged eastern segment of campylon plateau, located in pennyroyal and notte, which is a diversified, adjacent mixed middle forest of about 80 tree species, including northern red oak (Quercus rubus L.), white oak (Quercus alba L.), yellow poplar (L irden tulipera L1), american cockayne (fagusgradoliahehrh), eastern ferrugreek (Tsuga adadensis) (L) and western mangrove (american maple), which includes the dominant trees such as cornflower forest, western forest, northern mangrove ash, sanra, northern mangrove ash, northern oak, northern, southern, northern.
In summer of 2013, 271 round plots were field investigated over the entire area, and the area was 0.04 hectares (0.1ac) (one grid investigation per 384 meters). 271 plots were not georeferenced very accurately, with errors of up to 5 meters. Within each plot, all trees with a breast diameter greater than 12.5cm recorded species, breast diameter, height, crown grade (preponderance, sub-preponderance, intermediate or over-top), tree status (alive or dead), and trunk type (single or multiple). In addition, the horizontal distance and azimuth from the center of the plot to the surface of each tree at the breast diameter are collected to create a trunk map. The topographic features of various lands are recorded, including gradient, slope direction, slope position and the like. Table 1 shows the attribute statistics of the robinson forest 271 plots.
TABLE 1 Attribute statistics for 271 plots in RF
As shown in figure 1, the robust segmentation method for determining deciduous forest tree-level parameters by using the airborne laser radar comprises the following steps of scanning a low-altitude flying laser radar and a hollow flying laser radar in a deciduous forest region, respectively acquiring L eaf-on data and L eaf-off data, preprocessing and analyzing the L eaf-on data, then obtaining crown parameters by using a multi-threshold segmentation method, preprocessing and analyzing the L eaf-off data, separating all trunk radar points, clustering the obtained trunk radar points to obtain trunk parameters, matching the crown parameters with the trunk parameters to obtain single trees, finally segmenting and depicting the single trees, and extracting tree-level structure parameters from the single trees to complete the determination of the corresponding deciduous forest tree-level parameters.
It is noted that for L eaf-on data and L eaf-off data, which are a combination of two separate data sets collected by the same lidar system (Leica A L S60), where the L eaf-off data set is a low density (-2 pt/m2) data set collected in the 2013 spring season (fallen leaves season) for obtaining topographical attributes, which is part of the Kentucky department State elevation data set acquisition plan, the L eaf-on data set is high density data (-50 pt/m2) collected in the 2013 summer season (leafy flourishing season) for collecting detailed forest structure information.
TABLE 2 lidar data acquisition parameters for two data sets of RF
Further, the method of pre-processing the L eaf-on data or L eaf-off data is embodied by pre-processing two raw lidar point data sets (i.e., a leaf-off data set and a leaf-on data set) using Terrrascan software, dividing the lidar points into ground and non-ground points, L ASTools tool in ArcMAP 10.2 for generating a single airborne radar scan (L AS) data set file containing two L iDAR data sets, then generating a 1 meter resolution Digital Elevation Model (DEM) based on the ground points, filling in with nearest neighbors, interpolating using averaging, the pre-processing further comprises 4 steps of (1) normalizing the reach point spacing and creating a resolution grid equal to the Nominal Point Spacing (NPS), (2) selecting each grid cell AS the highest elevation point of the lidar surface points (L) to filter the lidar, (3) calculating all radar derived radar point spacing and using a smoothing filter 357, and a smoothing filter to create a few post-vegetation density variations in the ground tree (SP) and smoothing the top tree height using a smoothing filter L, and a smoothing filter to create a few post-vegetation density variation in the post-vegetation map analysis.
After the preprocessing procedure is completed, a multi-threshold tree segmentation method can be carried out on leaf-on data containing rich canopy information, and the trunk can be detected and separated from leaf-off point clouds in the same sample plot.
Further, as shown in fig. 2, the method for performing multi-threshold segmentation after preprocessing L eaf-on data includes the following steps:
(1) layering the point cloud according to the echo number of the laser radar;
(2) performing a multi-threshold segmentation within each separated canopy;
(3) combining crown sections in a cross-layer manner;
(4) the resulting individual crowns of the segmentation are depicted.
Since the multi-threshold segmentation method has been disclosed as the prior art, further description is omitted in this embodiment.
Unlike tree splitting based on a tree crown, noise points contained in a lidar cluster (radar points from shrubs or branches) have a large impact on the accuracy of trunk detection and identification. The clustering technology based on DBSCAN is a clustering method based on density, and can effectively reduce the influence of noise on clustering precision.
DBSCAN is a typical density-based clustering algorithm that can partition areas with sufficiently high density into clusters and find clusters of arbitrary shape in a spatial database containing noise. It describes the distribution of samples based on a set of "neighborhood" parameters (Eps, MinPts).
As shown in fig. 3, to introduce the DBSCAN clustering algorithm, a data set D ═ { x is given1，x2，…，xmDefine the following concepts:
eps-neighborhood for xj ∈ D, its Eps-neighborhood contains the sum of x in sample set DjSamples having a distance of not more than Eps, i.e. NEps(xj)＝{xi∈D|dist(xi,xj) Eps ≦ and the number of this set of subsamples is denoted | NEps(xj)|；
Core object: if xjThe Eps-neighborhood of (A) contains at least MinPts samples, i.e. | NEps(xj) | is not less than MinPts, then xjIs a core object;
the direct density can reach: if xjAt xiIn the Eps-field of (c), and xiIs a core object, then called xjFrom xiThe density is direct; note that the opposite does not necessarily hold, i.e., x cannot be said at this timeiFrom xjDensity is not direct unless xjIs also a core object;
the density can reach: for xiAnd xjIf a sample sequence p is present1，p2，...，pnWherein p is1＝xi，pn＝xjAnd p isi+1From piWhen the density is up to, it is called xjFrom xiThe density can be reached; at this point in the sequence the transfer sample p1，p2，...，pnAll the core objects are core objects, because only the core objects can enable the density of other samples to reach directly, the density can reach and also cannot meet the symmetry, and the density can be obtained from the asymmetry of the direct density;
density connection: for xiAnd xjIf x is presentkSo that xiAnd xjAre all xkWhen the density is up, it is called xiAnd xjConnecting the densities; note that the density connectivity is such that symmetry is satisfied.
Notably, DBSCAN requires a set of "neighborhood" parameters (Eps, MinPts) that need not be provided simultaneously. Eps represents the cluster radius, and Minpts represents the number of points on the circle, i.e., the density threshold of the cluster. When the point cloud shows a relatively uniform distribution, Eps may be calculated from the size of the point cloud and MinPts using the following formula:
where m denotes the number of points, n denotes the number of dimensions of the points, and T denotes the volume of the clustering space composed of m points.
On the basis of the above, the specific method for separating all the trunk radar points and dividing the trunk parts (as shown in fig. 4-a and 4-b) after preprocessing the L eaf-off data in the embodiment is as follows:
(1) splitting a lidar point cloud in a sample plot into N with a height of 30 meterslA layer;
(2) calculating laser radar point n of each layeriThe number of laser radar points and the percentage of the total number of laser radar points in each layer in the sample plot;
(3) forming a histogram in which points are distributed in the height direction;
(4) the density of search points exceeds a predetermined threshold hbaseThe lowest level layer l ofx；
(5) Will lxIs defined as a dividing plane,/xThe following points are then potential trunk radar points, including one or more trunks.
Further, the method for clustering the trunk radar points to obtain the trunk parameters is to cluster the trunk radar points according to their spatial neighborhood relationship to obtain the estimated trunk number and position in one sample plot, and allocate the trunk radar points to the trunk point cluster (as shown in fig. 5a-5 d), during this process, a trunk detection and identification method based on DBSCAN is also used to eliminate the points on bush or isolated branches in the trunk radar points, and to depict trunks of different shapes (such as bending and twisting), specifically, the method clusters the radar point areas with sufficiently high density into clusters, and searches for clusters of any shape in the noisy spatial database. Further, these radar points are assigned to the estimated trunk without assuming a trunk shape or number in advance. It describes the distribution of samples based on a set of "neighborhood" parameters (Eps, MinPts). Before performing cluster analysis, the stem points are projected onto a horizontal plane, and then a suitable set of parameters (Eps, MinPts) is used to implement a clustering process based on DBSCAN in a two-dimensional space formed by the projected points. Eps represents the cluster radius, and MinPts represents the number of points on the circle, i.e., the density threshold of the cluster. Initial values of the two parameters are calculated by the equations (1) and (2), and then the optimal values in the sample plot are obtained through multiple iterative comparisons. In the study case of this example, the ideal values are 2 and 5, respectively.
On the basis, further, data post-processing can be performed on the trunk point cluster, and the output is further improved by excluding all the following point clusters: a cluster of points with a minimum height greater than 4m (to be considered as a segment from the canopy point segment), a cluster of points with a maximum height less than 2m (to be considered as low vegetation), a cluster of trunk points with a quartering distance of less than 1m (to be considered as a fallen tree).
In this embodiment, the method for obtaining the single tree by matching the crown parameters and the trunk parameters includes the following steps:
(1) given forest canopy number ilay( i lay1, 2, 3), the detected trunk number is jstem(jstem＝0，1，2，…)；
(2) Projecting the segmented crown point clusters in each layer onto an x-y plane, and checking a trunk extraction result related to the crown point segments;
(3) depending on the detected number of trunks, the matching method of crown segments to trunks includes the following cases (as shown in fig. 6a-6 b):
(3.1) when the projected point of the trunk is not contained in the crown part in a certain projected plane (as R3 and R4 in fig. 6 a), i.e. the crown part will be regarded as a branch and merged to the tree closest to it;
(3.2) when a certain projected layer appears to have a trunk contained within a crown segment (e.g., R1, R2, R5, R6, and R7 in fig. 6 a), this means that the trunk matches a crown segment within a given crown layer, making the trunk and the crown segment within the given crown layer as candidate matching pairs, and then further determining whether their point clusters are combined together to outline the entire structure of the singles;
(3.3) when a certain projection surface layer appears that two or more trunks are contained in a crown section (such as R8 in FIG. 6 a), this situation means that there may be some small trees completely within the vertical projection range of a large tree crown and considered as a part of the large tree, so that it is judged that a plurality of small trees are completely within the vertical projection range of a large tree crown;
(3.4) when a case where no crown section is present in the region directly above the trunk of a certain projected course, it is judged that no crown section belongs to it, and it is regarded as a dry tree or a low shrub and deleted from the trunk point cluster.
Here, the tree crown segment cross-layer combination rule in the multi-threshold segmentation method can be improved to obtain the matching rule as shown in fig. 6-a. In fig. 6-a, after stem detection, two stems (stem 2 and stem 3) are contained in crown segment R8. According to the matching technique proposed in this embodiment (as shown in fig. 7), the crown segments R2 and R6 are combined into a tree and no longer merged into R8 as branches. The modified tree structure shown in fig. 6-b, i.e. the modification of the so-called "junction growing tree" (CGT) in the prior art multi-threshold segmentation method of patent No. 2019103511285, shows a combination of nodes within different canopies, each node representing a junction region, according to the above analysis method.
After selecting a set of pairs with the smallest distance, wherein the occurrence frequency of the trunk or the crown section does not exceed one time, the trunk is considered to be matched with the crown section. The matching crown segments are selected as root nodes (e.g., R5, R6, and R8 in fig. 6-a and 6-b) in this embodiment, which are not merged into other nodes. The remaining crown segment point clusters constitute the final segmentation result. 8-a, 8-b, 8-c, 8-d (where the different shapes of the lidar points represent different trees and R1, R2, and R3 represent tree segments for each canopy) show a comparison between the multi-threshold tree splitting method and the proposed trunk detection assisted multi-threshold tree splitting method (i.e., the method provided by the present embodiment). It can be seen that the tree that is not split (the tree marked with triangles in fig. 8-a) is accurately separated and depicted using the trunk detection assisted multi-threshold segmentation method (i.e., the method provided by the present embodiment).
Finally, the method for segmenting and describing the single trees and extracting tree-level structure parameters from the single trees comprises the steps that after the segmented single trees are determined through matching of the crown section point clusters and the trunk point clusters, the tree height, the crown size, the trunk position and the number of trees in the sample plot in the tree structure parameters can be directly derived from the segmented point clusters. When structural parameters of the tree are derived, the highest point in the point cluster of each laser radar derived tree is regarded as the top point of the tree, the vertical distance from the highest point to the ground is regarded as the height of the tree, and the perimeter of the tree crown is defined as the length of a boundary line of the tree crown projected onto an x-y plane.
In addition, this embodiment also provides an evaluation of the method of the present invention, and specifically, in order to verify the three-dimensional segmentation accuracy of a single Tree, this embodiment matches the estimation Tree (or lidar-derived Tree) and the reference Tree (or field measurement Tree) using a semi-automatic method proposed by Zhao et al (Zhao, k.; Suarez, j.c.; Garcia, m.; Hu, t.; Wang, c.; L ondo, a.utility of multiple lidar for purposes and carbon monitoring ordering: Tree growth, bioglass dynamics, and carbon flux.remote.sens.environ.2018,204, 883-897).
First, for two data sets (i.e., an estimate tree and a reference tree), two locations (x) are computed using a particular distance metric1，y1，z1) And (x)2，y2，z2) The distance between:
wherein (x)1，y1) And z1Respectively representing the trunk position and the tree height of the estimated tree; (x)2，y2) And z2Respectively representing the position and height of the trunk measured on site; as a user-defined parameter, w weights the vertical distance difference with the planar distance (in this study, the value of w is 0.5).
Second, the location of a tree in one dataset is paired into another dataset only if the two trees are located closest to each other. The preliminary list of paired trees is automatically obtained by running the above matching method using the distance metric of equation (5).
And finally, selecting a pairing set which has the minimum distance and the estimated or reference tree position does not occur more than once by using a Hungarian allocation algorithm, and regarding the paired data set as a matching tree.
For tree level evaluation, if a tree actually exists and can be segmented from the radar point cloud, it is called a Matching Tree (MT); if a tree is not split up but is assigned to other trees as branches, it is called an miss errors (OE) tree; if a tree does not exist but is segmented from the radar point cloud, it is called a Commission Error (CE) tree. The number of OE and CE trees represents under-segmentation and over-segmentation, respectively. To evaluate the accuracy of this method, this example calculates the Recall (Recall, Re), Precision (Precision, Pr), and F score (F-score) using the following formulas:
where Re is a measure of the detection rate of the tree, Pr is a measure of the correctness of the detected tree, and F is a measure of the overall accuracy after considering CE and OE.
Statistics of tree splitting using the method provided in this embodiment (which may also be referred to as "trunk detection assisted multi-threshold detection method"), multi-threshold method, and DSM-based method are shown in table 3. The trunk detection assisted multi-threshold method proposed herein detects more trees in 271 plots than the latter. Compared with a multi-threshold method and a method based on DSM, the average detection rate is respectively improved by 4.5 percentage points and 17.5 percentage points, and the correct rate of the detection tree is respectively improved by 3.1 percentage points and 3.4 percentage points. Accordingly, the overall accuracy of the segmentation improved by 4.6 percentage points and 12.1 percentage points, respectively, considering CE and OE.
Table 3. the number of segmented trees obtained by applying the method proposed in this example, the multi-threshold method, and the DSM method to 271 sample plots (3896 measured trees), and statistics of Re, Pr, and F, respectively.
The statistical results of the division scores of the upper and lower forest trees of the 271 plots are shown in table 4. Compared with the multi-threshold method and the DSM method, for the upper forest, Re is respectively improved by 0.5 percentage point and 3.1 percentage point, Pr is respectively improved by 2.7 percentage point and 4.1 percentage point, and F is respectively improved by 2.2 percentage point and 4.6 percentage point. Re of the lower forest is respectively improved by 2.0 percent and 23.1 percent, Pr is respectively improved by 3.6 percent and 1.8 percent, and F is respectively improved by 4.8 percent and 16.5 percent.
Table 4.271 plots of upper and lower trees.
In addition, the present embodiment also evaluates the performance of tree-level parameter estimation after matching the estimation tree with the reference tree. Two tree structure parameters (tree position and height) obtained using lidar data were verified against the measured tree position and height. Only correctly detected trees are used; that is, the erroneously detected trees and the lost trees are not calculated. For field survey data, of 271 plots of 1240 total trees, 84 plots were measured for tree height. These 84 blocks are similarly subjected to the proposed method, multi-threshold segmentation method and DSM-based method to detect 1159, 1135 and 897 trees, respectively. The relationship between the height of all correctly detected estimated trees and the reference tree obtained by the three methods is shown in fig. 9. The linear regression fits (R2) for the reference tree height and the estimated tree height were 0.96, 0.95, and 0.90, respectively.
After collecting the trunk map of the radar-derived tree, the distance distribution between all correctly detected radar-derived trees and reference tree locations is obtained, as shown in fig. 10, as previously described, both the multi-threshold segmentation and DSM-based methods utilize the lidar surface points (L SP) of the canopy (or canopy layers of different heights) to identify the contours of the segmented trees, which is a canopy-based segmentation method.
Through the evaluation and verification, the robust segmentation method for determining the deciduous forest tree-level parameters by adopting the airborne laser radar, which is provided by the invention, utilizes the additional information in the tree, and obtains better segmentation effect compared with a DSM-based method and a multi-threshold method previously proposed by the applicant. The method not only improves the detection rate of the tree, but also improves the correctness of the detected tree, can more accurately estimate tree-level structure parameters such as the tree height, the tree crown perimeter, the trunk position, the tree crown and the like, is favorable for the rapid development of the single-tree-based method in the aspect of airborne laser radar data application, and has important significance in improving the accuracy and the efficiency direction of forest resource clearing.
The above embodiments are preferred implementations of the present invention, and the present invention can be implemented in other ways without departing from the spirit of the present invention.
Claims (10)
1. The robust segmentation method for determining deciduous forest tree-level parameters by adopting an airborne laser radar is characterized in that low-altitude flight laser radar scanning and hollow flight laser radar scanning are carried out on a deciduous forest region, L eaf-on data and L eaf-off data are acquired respectively, a multi-threshold segmentation method is adopted after preprocessing and analyzing the L eaf-on data to obtain crown parameters, all trunk radar points are separated after preprocessing and analyzing the L eaf-off data, then the obtained trunk radar points are clustered to obtain trunk parameters, the crown parameters and the trunk parameters are matched to obtain single trees, finally the single trees are segmented and depicted, and tree-level structure parameters are extracted from the single trees to complete determination of the corresponding deciduous forest tree-level parameters.
2. The robust segmentation method for deciduous forest tree-level parameterization using an airborne lidar according to claim 1, wherein the method of pre-processing the L eaf-on data or L eaf-off data comprises the steps of:
(1) dividing laser radar points in corresponding data into ground points and non-ground points, generating a Digital Elevation Model (DEM) with the resolution of 1 meter based on the ground points, performing gap filling by adopting a nearest neighbor method, and performing interpolation by adopting an average method;
(2) normalizing radar spot spacing and creating a resolution grid equal to a nominal spot spacing (NPS);
(3) selecting each grid cell as a highest elevation point of a lidar surface point (L SP) to filter the lidar point cloud;
(4) calculating the ground level of all L SPs using a lidar derived DEM;
(5) the L SP was smoothed using a gaussian smoothing filter with a standard deviation equal to NPS and a radius of 3 × NPS to reduce small variations in canopy implant height.
3. The robust segmentation method for deciduous forest tree-level parameter measurement by using the airborne lidar as claimed in claim 2, wherein the method for performing multi-threshold segmentation after preprocessing the L eaf-on data comprises the following steps:
(1) layering the point cloud according to the echo number of the laser radar;
(2) performing a multi-threshold segmentation within each separated canopy;
(3) combining crown sections in a cross-layer manner;
(4) the resulting individual crowns of the segmentation are depicted.
4. The robust segmentation method for deciduous forest tree-level parameter measurement by using the airborne laser radar as claimed in claim 3, wherein the method for separating all trunk radar points after preprocessing the L eaf-off data comprises the following steps:
(1) splitting a lidar point cloud in a sample plot into N with a height of 30 meterslA layer;
(2) calculating laser radar point n of each layeriThe number of laser radar points and the percentage of the total number of laser radar points in each layer in the sample plot;
(3) forming a histogram in which points are distributed in the height direction;
(4) the density of search points exceeds a predetermined threshold hbaseThe lowest level layer l ofx；
(5) Will lxIs defined as a dividing plane,/xThe following points are then potential trunk radar points, including one or more trunks.
5. The robust segmentation method for deciduous forest tree-level parametric measurement using airborne lidar according to claim 4, wherein: the method for clustering the trunk radar points to obtain the trunk parameters comprises the steps of clustering the trunk radar points according to the spatial neighborhood relationship of the trunk radar points to obtain the estimated trunk number and position in a sample plot, and distributing the trunk radar points to the trunk point cluster.
6. The robust segmentation method for deciduous forest tree-level parametric measurement using airborne lidar according to claim 5, wherein: the trunk detection and identification method based on the DBSCAN is adopted to eliminate the points on shrubs or isolated branches in the trunk radar points and draw trunks with different shapes, wherein the trunk detection and identification method based on the DBSCAN comprises the following steps:
(1) assigning the trunk radar points to estimated trunks without presupposing trunk shapes or numbers, wherein the trunk points are all projected onto a horizontal plane;
(2) the clustering process based on the DBSCAN is realized in a two-dimensional space formed by projection points on a horizontal plane of the previous step by using a group of parameters (Eps, MinPts), wherein Eps represents a clustering radius, MinPts represents the number of points on a circle, namely a clustering density threshold, the initial values of the two parameters are calculated by the formula (1) and the formula (2), and then the optimal value in a sample plot is obtained through multiple iterative comparisons;
7. the robust segmentation method for deciduous forest tree-level parametric measurement using airborne lidar according to claim 6, wherein: and performing data post-processing on the trunk point cluster, and excluding the point clusters with the minimum height larger than 4m, the point clusters with the maximum height smaller than 2m and the trunk point clusters with the four-division distances smaller than 1m to finally obtain trunk parameters.
8. The robust segmentation method for deciduous forest tree-level parametric measurement using airborne lidar according to claim 7, wherein: the method for obtaining the single tree by matching the crown parameters with the trunk parameters comprises the following steps:
(1) given forest canopy number ilay(ilay1, 2, 3), the detected trunk number is jstem(jstem＝0，1，2，…)；
(2) Projecting the segmented crown point clusters in each layer onto an x-y plane, and checking a trunk extraction result related to the crown point segments;
(3) according to the detected number of the trunks, the matching method of the crown sections and the trunks comprises the following conditions:
(3.1) when the projected point of the trunk is not contained in the crown part in a certain projection plane, namely the crown part is regarded as a branch and is merged to the tree closest to the branch;
(3.2) when a trunk appears in a certain projection layer to be contained in a crown section, enabling the trunk and the crown section in the given crown layer to be used as a candidate matching pair, and further determining whether the point clusters are combined together to outline the whole structure of the single tree;
(3.3) when two or more trunks are contained in one crown section in a certain projection surface layer, judging that the plurality of small trees are completely in the vertical projection range of the crown of one large tree;
(3.4) when a case where no crown section is present in the region directly above the trunk of a certain projected course, it is judged that no crown section belongs to it, and it is regarded as a dry tree or a low shrub and deleted from the trunk point cluster.
9. The robust segmentation method for deciduous forest tree-level parametric measurement using airborne lidar according to claim 8, wherein: the method for segmenting and describing the single tree and extracting tree-level structure parameters from the single tree comprises the steps that after the segmented single tree is determined through matching of the crown section point cluster and the trunk point cluster, the tree height, the crown size, the trunk position and the number of trees in a sample plot in the tree structure parameters can be directly derived from the segmented point clusters.
10. The robust segmentation method for deciduous forest tree-level parametric measurement using airborne lidar according to claim 9, wherein: when structural parameters of the tree are derived, the highest point in the point cluster of each laser radar derived tree is regarded as the top point of the tree, the vertical distance from the highest point to the ground is regarded as the height of the tree, and the perimeter of the tree crown is defined as the length of a boundary line of the tree crown projected onto an x-y plane.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010206026.7A CN111428784A (en) | 2020-03-23 | 2020-03-23 | Robust segmentation method for deciduous forest tree-level parameter measurement by adopting airborne laser radar |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010206026.7A CN111428784A (en) | 2020-03-23 | 2020-03-23 | Robust segmentation method for deciduous forest tree-level parameter measurement by adopting airborne laser radar |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN111428784A true CN111428784A (en) | 2020-07-17 |
Family
ID=71553637
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010206026.7A Pending CN111428784A (en) | 2020-03-23 | 2020-03-23 | Robust segmentation method for deciduous forest tree-level parameter measurement by adopting airborne laser radar |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111428784A (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112837309A (en) * | 2021-03-02 | 2021-05-25 | 华南农业大学 | Fruit tree canopy target recognition device and method, computing equipment and storage medium |
| CN113222917A (en) * | 2021-04-29 | 2021-08-06 | 桂林理工大学 | DBI tree vertex detection method of airborne laser radar point cloud data CHM |
| CN113449733A (en) * | 2021-03-24 | 2021-09-28 | 西安理工大学 | Point cloud scene tree extraction method based on structural analysis |
| CN113591766A (en) * | 2021-08-09 | 2021-11-02 | 中国林业科学研究院资源信息研究所 | Multi-source remote sensing tree species identification method for unmanned aerial vehicle |
| WO2023085996A1 (en) * | 2021-11-09 | 2023-05-19 | Saab Ab | Method, system and computer program product for forming a digital surface model based on treetops |
-
2020
- 2020-03-23 CN CN202010206026.7A patent/CN111428784A/en active Pending
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112837309A (en) * | 2021-03-02 | 2021-05-25 | 华南农业大学 | Fruit tree canopy target recognition device and method, computing equipment and storage medium |
| CN112837309B (en) * | 2021-03-02 | 2023-10-20 | 华南农业大学 | Fruit tree canopy target recognition device, method, computing equipment and storage medium |
| CN113449733A (en) * | 2021-03-24 | 2021-09-28 | 西安理工大学 | Point cloud scene tree extraction method based on structural analysis |
| CN113449733B (en) * | 2021-03-24 | 2023-04-18 | 西安理工大学 | Point cloud scene tree extraction method based on structural analysis |
| CN113222917A (en) * | 2021-04-29 | 2021-08-06 | 桂林理工大学 | DBI tree vertex detection method of airborne laser radar point cloud data CHM |
| CN113591766A (en) * | 2021-08-09 | 2021-11-02 | 中国林业科学研究院资源信息研究所 | Multi-source remote sensing tree species identification method for unmanned aerial vehicle |
| WO2023085996A1 (en) * | 2021-11-09 | 2023-05-19 | Saab Ab | Method, system and computer program product for forming a digital surface model based on treetops |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111428784A (en) | Robust segmentation method for deciduous forest tree-level parameter measurement by adopting airborne laser radar | |
| CN106815847B (en) | Trees dividing method and single tree extracting method based on laser radar point cloud | |
| CN107832681B (en) | Forest single tree high estimation method combining LiDAR point cloud and synchronous remote sensing image | |
| Li et al. | A dual growing method for the automatic extraction of individual trees from mobile laser scanning data | |
| Lefsky et al. | Volume estimates of trees with complex architecture from terrestrial laser scanning | |
| CN106529469A (en) | Unmanned aerial vehicle airborne LiDAR point cloud filtering method based on adaptive gradient | |
| CN106407925B (en) | Laser scanning point cloud trees extraction method based on local section maximum | |
| CN105488770A (en) | Object-oriented airborne laser radar point cloud filtering method | |
| CN112381861A (en) | Forest land point cloud data registration and segmentation method based on foundation laser radar | |
| CN112347894B (en) | Single plant vegetation extraction method based on transfer learning and Gaussian mixture model separation | |
| CN113269825B (en) | Forest breast diameter value extraction method based on foundation laser radar technology | |
| CN109961470B (en) | Living standing tree She Shuxing accurate estimation method based on laser point cloud | |
| Pirotti et al. | A comparison of tree segmentation methods using very high density airborne laser scanner data | |
| CN114241217B (en) | Trunk point cloud efficient extraction method based on cylindrical features | |
| CN111950589B (en) | Point cloud region growing optimization segmentation method combined with K-means clustering | |
| CN110196432B (en) | Deciduous forest tree-level parameter determination method based on small-spot airborne radar | |
| CN113780144A (en) | Crop plant number and stem width automatic extraction method based on 3D point cloud | |
| Wang et al. | Combining trunk detection with canopy segmentation to delineate single deciduous trees using airborne LiDAR data | |
| Smits et al. | Individual tree identification using different LIDAR and optical imagery data processing methods | |
| Lei et al. | Extraction of maize leaf base and inclination angles using terrestrial laser scanning (TLS) data | |
| Zhu et al. | A method for detecting tomato canopies’ phenotypic traits based on improved skeleton extraction algorithm | |
| Gatziolis et al. | Evaluation of terrestrial LiDAR technology for the development of local tree volume equations | |
| CN110348478B (en) | Method for extracting trees in outdoor point cloud scene based on shape classification and combination | |
| Zhu et al. | Research on deep learning individual tree segmentation method coupling RetinaNet and point cloud clustering | |
| Bortolot | Using tree clusters to derive forest properties from small footprint lidar data |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |