lidR (version 4.1.0)

segment: Segment a point cloud

Description

Segment a point cloud using different methods. segment_* functions add a new attribute to the point cloud to label each point. They segment either individual trees, snags, or geometrical features.

Usage

segment_shapes(las, algorithm, attribute = "Shape", filter = NULL)
segment_snags(las, algorithm, attribute = "snagCls")
segment_trees(las, algorithm, attribute = "treeID", uniqueness = "incremental")

Arguments

las

An object of class LAS or LAScatalog.

algorithm

function. An algorithm for segmentation. For individual tree segmentation, lidR has dalponte2016, watershed, li2012, and silva2016. More experimental algorithms may be found in the package lidRplugins. For snag segmentation, lidR has wing2015. For geometry segmentation, lidR has shp_plane, shp_hplane, and shp_line.

attribute

character. The returned LAS object as a new attribute (in a new column). This parameter controls the name of the new attribute.

filter

formula of logical predicates. Enables the function to run only on points of interest in an optimized way. See the examples.

uniqueness

character. A method to compute a unique ID. Can be 'incremental', 'gpstime' or 'bitmerge'. See section 'Uniqueness'. This feature must be considered as 'experimental'.

Non-supported LAScatalog options

The option select is not supported and not respected because it always preserves the file format and all the attributes. select = "*" is imposed internally.

Uniqueness

By default the tree IDs are numbered from 1 to n, n being the number of trees found. The problem with such incremental numbering is that, while it ensures a unique ID is assigned for each tree in a given point-cloud, it also guarantees duplication of tree IDs in different tiles or chunks when processing a LAScatalog. This is because each chunk/file is processed independently of the others and potentially in parallel on different computers. Thus, the index always restarts at 1 on each chunk/file. Worse, in a tree segmentation process, a tree that is located exactly between 2 chunks/files will have two different IDs for its two halves.

This is why we introduced some uniqueness strategies that are all imperfect and that should be seen as experimental. Please report any troubleshooting. Using a uniqueness-safe strategy ensures that trees from different files will not share the same IDs. It also ensures that two halves of a tree on the edge of a processing chunk will be assigned the same ID.

incremental

Number from 0 to n. This method does not ensure uniqueness of the IDs. This is the legacy method.

gpstime

This method uses the gpstime of the highest point of a tree (apex) to create a unique ID. This ID is not an integer but a 64-bit decimal number, which is suboptimal but at least it is expected to be unique if the gpstime attribute is consistent across files. If inconsistencies with gpstime are reported (for example gpstime records the week time and was reset to 0 in a coverage that takes more than a week to complete), there is a (low) probability of getting ID attribution errors.

bitmerge

This method uses the XY coordinates of the highest point (apex) of a tree to create a single 64-bit number with a bitwise operation. First, XY coordinates are converted to 32-bit integers using the scales and offsets of the point cloud. For example, if the apex is at (10.32, 25.64) with a scale factor of 0.01 and an offset of 0, the 32-bit integer coordinates are X = 1032 and Y = 2564. Their binary representations are, respectively, (here displayed as 16 bits) 0000010000001000 and 0000101000000100. X is shifted by 32 bits and becomes a 64-bit integer. Y is kept as-is and the binary representations are unionized into a 64-bit integer like (here displayed as 32 bit) 00000100000010000000101000000100 that is guaranteed to be unique. However R does not support 64-bit integers. The previous steps are done at C++ level and the 64-bit binary representation is reinterpreted into a 64-bit decimal number to be returned in R. The IDs thus generated are somewhat weird. For example, the tree ID 00000100000010000000101000000100 which is 67635716 if interpreted as an integer becomes 3.34164837074751323479078607289E-316 if interpreted as a decimal number. This is far from optimal but at least it is guaranteed to be unique if all files have the same offsets and scale factors.

All the proposed options are suboptimal because they either do not guarantee uniqueness in all cases (inconsistencies in the collection of files), or they imply that IDs are based on non-integers or meaningless numbers. But at least it works and deals with some of the limitations of R.

Details

segment_trees

Individual tree segmentation with several possible algorithms. The returned point cloud has a new extra byte attribute named after the parameter attribute independently of the algorithm used.

segment_shapes

Computes, for each point, the eigenvalues of the covariance matrix of the neighbouring points. The eigenvalues are later used either to segment linear/planar points or to compute derived metrics. The points that meet a given criterion based on the eigenvalue are labelled as approximately coplanar/colinear or any other shape supported.

segment_snags

Snag segmentation using several possible algorithms. The function attributes a number identifying a snag class (snagCls attribute) to each point of the point cloud. The classification/segmentation is done at the point cloud level and currently only one algorithm is implemented, which uses LiDAR intensity thresholds and specified neighbourhoods to differentiate bole and branch from foliage points.

Examples

# ==============
# Segment trees
# ==============

LASfile <- system.file("extdata", "MixedConifer.laz", package="lidR")
las <- readLAS(LASfile, select = "xyz", filter = "-drop_z_below 0")

# Using Li et al. (2012)
las <- segment_trees(las, li2012(R = 3, speed_up = 5))
#plot(las, color = "treeID")

# ==============
# Segment shapes
# ==============

LASfile <- system.file("extdata", "Megaplot.laz", package="lidR")
las <- readLAS(LASfile, filter = "-keep_random_fraction 0.5")

# Use the eigenvalues to estimate if points are part of a local plan
las <- segment_shapes(las, shp_plane(k = 15), "Coplanar")
#plot(las, color = "Coplanar")

if (FALSE) {
# Drop ground point at runtime
las <- segment_shapes(las, shp_plane(k = 15), "Coplanar", filter = ~Classification != 2L)
#plot(las, color = "Coplanar")

# ==============
# Segment snags
# ==============

LASfile <- system.file("extdata", "MixedConifer.laz", package="lidR")
las <- readLAS(LASfile, select = "xyzi", filter="-keep_first") # Wing also included -keep_single

# For the Wing2015 method, supply a matrix of snag BranchBolePtRatio conditional
# assessment thresholds (see Wing et al. 2015, Table 2, pg. 172)
bbpr_thresholds <- matrix(c(0.80, 0.80, 0.70,
                            0.85, 0.85, 0.60,
                            0.80, 0.80, 0.60,
                            0.90, 0.90, 0.55),
                            nrow =3, ncol = 4)

# Run snag classification and assign classes to each point
las <- segment_snags(las, wing2015(neigh_radii = c(1.5, 1, 2), BBPRthrsh_mat = bbpr_thresholds))

# Plot it all, tree and snag points...
plot(las, color="snagCls", colorPalette = rainbow(5))

# Filter and plot snag points only
snags <- filter_poi(las, snagCls > 0)
plot(snags, color="snagCls", colorPalette = rainbow(5)[-1])

# Wing et al's (2015) methods ended with performing tree segmentation on the
# classified and filtered point cloud using the watershed method
}