Airborne LiDAR data manipulation and visualisation for forestry application
unormalize_height()
removes extra_bytes in VLR.print(las)
works even when the CRS is not recognized by sf
.dsmtin
and pitfree
gain an argument highest
. This option was enabled by default in previous releases. There is now an option to disable it.normalize_height()
and segment_trees
work in parallel with SpatRaster
.crown_metrics()
now triggers a warning when invalid geometries are created and delineate_crowns()
remove these geometries before to convert to sp
.crown_metrics()
now works with func = NULL
and a LAScatalog
.*_metrics()
functions always returned NA
s for lastofmany
.dalponte2016
doc updated to use terra
.plot(ctg, chunk = TRUE)
does not fail if an invalid output file template is registered #537
locate_trees()
throws an informative error if called with an on-disk raster. The former error was cryptic. If the raster is small enough it is loaded on-the-fly.merge_spatial()
with RGB and SpatRaster
was not working properly #545
st_area()
better estimates the area of small point-clouds and is fasterinterpret_waveform
#549.plot_metrics()
returns NA if 0 points available #551.rasterize_canopy
may generate error or messed-up CHM #552.print()
and st_area()
were not working for point cloud with no CRStrack_sensor()
does not fail with a LAScatalog
when no sensor position is found. It also triggers a warning. #556.rasterize_terrain()
now works with a LAScatalog
and shape = sfc_object
#558.catalog_retile()
now works when some tiles are empty #563.crown_metrics()
messed up tree IDs with a hull geometry #554.merge_spatial()
crops large vectors to the extent of the point cloud before to perform the merge. This has for consequences to sometime transform polygons into multipolygons. When polygons and multipolygons were mixed the functions stopped with an error. It now works.normalize_height()
now sets the Z offset to 0 #571.We are currently developing rlas 1.6.0 that uses the ALTREP framework to load compact representation of non populated attributes. For example UserData
is usually populated with zeros (not populated). Yet it takes 32 bits per point to store each 0. With rlas 1.6.0 it will only uses 644 bits no matter the number of points loaded for non populated attributes. This applies to each attribute populated with a single repeated value. This allows for saving approximately 30% of memory usage depending on the number of non-populated attributes that are present in the file. rlas 1.6.0 is compatible will all versions of lidR but lidR 4.0.1 introduced some internal optimization, internal fixes and new functions to fully take advantage of rlas 1.6.0. lidR v<= 4.0.0 will work with rlas 1.6.0 but won't take advantage of the new compression feature.
the function LAS()
no longer call data.table::setDT()
if the input is already a data.table
. Indeed data.table::setDT()
materializes the compressed ALTREP vectors and this is not what we want. One consequence of this change is that readLAS()
now preserve the ALTREPness (i.e. the compression) of the output of rlas::read.las()
.
Subsetting a LAS
object no longer call data.table
native subset. We previously used something like las@data[indx]
to subset the point cloud. Sadly data.table
tries to materialized the ALTREPed vector whenever it can. We implemented internally a smart_subset()
function that subset and preserves the compression of the vectors. One consequence of such change is that all filter_*()
and clip_*()
functions preserve the compression of the point-cloud if any.
las_check()
has been slightly modified to ensure it does not materialize ALTREPed object. One side effect of las_check()
was to decompress the point cloud unexpectedly. Such a pity! We also change las_check()
to print information about the compression.
We changed the way *_metrics()
functions evaluates the user defined expression because we found that it had the side effect of materializing all the attributes instead of materializing only those needed. For example pixel_metrics(las, mean(Z))
only needs the attribute Z. No need to allocate and copy memory for Intensity
, ScanAngle
and so on. In previous version all attributes where inspected with the side effect to materialize all compressed vectors. The *_metrics()
functions now properly detect which attributes are actually necessary for the evaluation of func
. Two consequences: (1)*_metrics()
functions are 20 to 40% faster, (2) the compression is preserved if no compressed attribute is used in the evaluation and e.g. pixel_metrics(las, mean(UserData))
uncompresses only UserData
.
New functions las_is_compressed()
that tells which attributes are compressed and las_size()
that returns the true size of a LAS
objects taking into account the compression. las_size()
should returns something similar to pryr::object_size()
but different to object.size()
that is not ALTREP aware. We also changed the print
function so it uses las_size()
instead of object.size()
.
On overall lidR's functions are expected to almost never decompress a LAS object. However other R packages and R functions may do it. For example data.table::print
do materializes the ALTREP vectors. base::range()
too but not base::mean()
or base::var()
.
las@data # Full decompression (print data.table)
range(las$Userdata) # Decompression of UserData
las@data[2, UserData := 1] # Decompression of UserData
las@data[1:10] # Full decompression
rgdal
and rgeos
will be retired on Jan 1st 2024. see twitter, youtube, or see the respective package descriptions on CRAN. Packages raster
and sp
are based on rgdal
/rgeos
and lidR
was based on raster
and sp
because it was created before sf
, terra
and stars
. This means that sooner or later lidR
will run into trouble (actually it is more or less already the case). Consequently, we modernized lidR
by moving to sf
, terra
/stars
and we are no longer depending on sp
and raster
(see also Older R Spatial Package for more insight). It is time for everybody to stop using sp
and raster
and to embrace sf
and stars/terra
.
In version 4 lidR
now no longer uses sp
, it uses sf
and it no longer uses raster
. It is now raster agnostic and works transparently with rasters from raster
, terra
and stars
. These two changes meant we had to rewrite a large portion of the code base, which implies few backward incompatibilities. The backward incompatibilities are very small compared to the huge internal changes we implemented in the foundations of the code and should not even be visible for most users.
lidR
no longer loads raster
and sp
. To manipulate Raster*
and Spatial*
objects returned by lidR users need to load sp
and raster
with:
library(sp)
library(raster)
library(lidR)
The formal class LAS
no longer inherits the class Spatial
from sp
. It means, among other things, that a LAS
object no longer has a slot @proj4string
with a CRS
from sp
, or a slot @bbox
. The CRS is now stored in the slot @crs
in a crs
object from sf
. Former functions crs()
and projection()
inherited from raster
are backward compatible and return a CRS
or a proj4string
from sp
. However code that accesses these slots manually are no longer valid (but nobody was supposed to do that anyway because it was the purpose of the function projection()
):
las@proj4string # No longer works
las@bbox # No longer works
inherits(las, "Spatial") # Now returns FALSE
The formal class LAScatalog
no longer inherits the class SpatialPolygonDataFrame
from sp
. It means, among other things, that a LAScatalog
object no longer has a slot @proj4string
, or @bbox
, or @polygons
. The slot @data
is preserved and contains an sf,data.frame
instead of a data.frame
allowing backward compatibility of data access to be maintained. The syntax ctg$attribute
is the way to access data, but statement like ctg@data$attribute
are backward compatible. However, code that accesses other slots manually is no longer valid, like for the LAS
class:
ctg@proj4string # No longer works
ctg@bbox # No longer works
ctg@polygons # No longer works
inherits(ctg, "Spatial") # Now returns FALSE
sp::spplot()
no longer works on a LAScatalog
because a LAScatalog
is no longer a SpatialPolygonDataFrame
spplot(ctg, "Max.Z")
# becomes
plot(ctg["Max.Z"])
raster::projection()
no longer works on LAS*
objects because they no longer inherit Spatial
. Moreover, lidR
no longer Depends
on raster
which means that raster::projection()
and lidR::projection
can mask each other. Users should use st_crs()
preferentially. To use projection
users can either load raster
before lidR
or call lidR::projection()
with the explicit namespace.
library(lidR)
projection(las) # works
library(raster)
projection(las) # no longer works
Serialized LAS/LAScatalog
objects (i.e. stored in .rds
or .Rdata
files) saved with lidR v3.x.y
are no longer compatible with lidR v4.x.y
. Indeed, the structure of a LAS/LAScatalog
object is now different mainly because the slot @crs
replaces the slot @proj4string
. Users may get errors when using e.g. readRDS(las.rds)
to load back an R object. However we put safeguards in place so, in practice, it should be backward compatible transparently, and even repaired automatically in some circumstances. Consequently we are not sure it is a backward incompatibility because we handled and fixed all warnings and errors we found. In the worst case it is possible to repair a LAS
object v3 with:
las <- LAS(las)
track_sensor()
is not backward compatible because it is a very specific function used by probably just 10 people in the world. We chose not to rename it. It now returns an sf
object instead of a SpatialPointsDataFrame
.
Former functions that return Spatial*
objects from package sp
should no longer be used. It is time for everybody to embrace sf
. However, these functions are still in lidR
for backward compatibility. They won't be removed except if package sp
is removed from CRAN. It might happen on Jan 1st 2024, it might happen later. We do not know. New functions return sf
or sfc
objects. Old functions are not documented so new users won't be able to use them.
tree_metrics()
and delineate_crowns()
are replaced by a single function crown_metrics()
that has the same functionality, and more.find_trees()
is replaced by locate_trees()
.Older functions that return Raster*
objects from the raster
package should no longer be used. It is time for everybody to embrace terra/stars
. However, these functions are still in lidR
for backward compatibility. They won't be removed except if package raster
is removed from CRAN. New functions return either a Raster*
, a SpatRaster
, or a stars
object, according to user preference.
grid_metrics()
is replaced by pixel_metrics()
grid_terrain()
, grid_canopy()
, grid_density()
are replaced by rasterize_terrain()
, rasterize_canopy()
, rasterize_density()
New functions are mostly convenient features that simplify some workflow aspects without introducing a lot of brand new functionality that did not already exist in lidR
v3.
New geometry functions st_convex_hull()
and st_concave_hull()
that return sfc
New modern functions st_area()
, st_bbox()
, st_transform()
and st_crs()
inherited from sf
for LAS*
objects.
New convenient functions nrow()
, ncol()
, dim()
, names()
inherited from base
for LAS*
objects
New operators $
, [[
, $<-
and [[<-
on LASheader
. The following are now valid statements:
header[["Version Major"]]
header[["Z scale factor"]] <- 0.001
Operators $
, [[
, $<-
and [[<-
on LAS
can now access the LASheader
metadata. The following are now valid statements:
las[["Version Major"]]
las[["Z scale factor"]] <- 0.001
RStudio now supports auto completion for operator $
in LAS
objects. Yay!
New functions template_metrics()
, hexagon_metrics()
, polygon_metrics()
that extend the concept of metrics further to any kind of template.
Functions that used to accept spatial vector or spatial raster as input now consistently accept any of Spatial*
, sf
, sfc
, Raster*
, SpatRaster
and stars
objects. This include merge_spatial()
, normalize_intensity()
, normalize_height()
, rasterize_*()
, segment_trees()
, plot_dtm3d()
and several others. We plan to support SpatVector
in future releases.
Every function that supports a raster as input now accept an "on-disk" raster from raster
, terra
and stars
i.e. a raster not loaded in memory. This includes rasterization functions, individual tree segmentation functions, merge_spatial
and others, in particular plot_dtm3d()
and add_dtm3d()
that now downsample on-disk rasters on-the-fly to display very large DTMs. On-disk rasters were already generally supported in previous versions but not every function was properly optimized to handle such objects.
All the functions that return a raster (pixel_metrics()
and rasterize_*()
) are raster agnostic and can return rasters from raster
, terra
or stars
. They have an argument pkg = "raster|terra|stars"
to choose. The default is terra
but this can be changed globally using:
options(lidR.raster.default = "stars")
New function catalog_map()
that simplifies catalog_apply()
to a large degree. Yet it is not as versatile as catalog_apply()
but well suits around 80% of use cases. Applying a user-defined function to a collection of LAS files is now as simple as:
my_fun <- function(las, ...) {
# do something with the point cloud
return(something)
}
res <- catalog_map(ctg, my_fun, param1 = 2, param2 = 5)
Operator [
on LAS
object has been overloaded to clip a point-cloud using a bbox
or a sfc
sub <- las[sfc]
rasterize_terrain()
accepts an sfc
as argument to force interpolation within a defined area.
normalize_height()
now always interpolates all points. It is no longer possible to get an error that some points cannot be interpolated. The problem of interpolating the DTM where there is no data is still present but we opted for a nearest neighbour approach with a warning instead of a failure. This prevents the method from failing after hours of computation for special cases somewhere in the file collection. This also means we removed the na.rm
option that is no longer relevant.
New functions header()
, payload()
, phb()
, vlr()
, evlr()
to get the corresponding data from a LAS
object.
New algorithm shp_hline
and shp_vline
for segment_shapes()
#499
New algorithm mcc
for ground classification.
The bounding box of the CHM computed with rastertize_canopy()
or grid_canopy()
is no longer affected by the subcircle
tweak. See #518.
readLAS()
can now read two or more files that do not have the same point format (see #508)
plot()
for LAS
gains arguments pal
, breaks
and nbreaks
similar to sf
. Arguments trim
and colorPalette
are deprecated
itot
from stdmetrics_i
which generates troubles (see #463 #514) is now double
instead of int
classify_*
, rasterize_*
, *_metrics
, segment_*
and normalize_*
were grouped.grid_*()
functions support a RasterLayer
smaller than the point cloud (#483)las_check()
with a LAScatalog
and with deep = TRUE
failed with a output file template (#484).readLAS()
no longer reads LAS files on some Windows/Mac machine (#485). It seems it is an issue with CRAN binaries. By releasing 3.2.2 we hope to trigger a new build.get_range()
and consequently range_correction()
no longer throw high range error for highly variable range sensor like TLS (#490).rgdal
and rgeos
will be retired on Jan 1st 2024. raster
and sp
are based on rgdal
/rgeos
. lidR
is based on raster
and sp
because it was created before sf
, terra
and stars
. This means that sooner or later lidR
will run into trouble (actually it has already started to be the case). So, it is time to fully embrace sf
, terra
/stars
and to leave sp
and raster
. This will require an in-depth rebase of lidR
. We have started the work and we plan to release lidR
4.0.0 that will no longer have any internal code that uses sp
and raster
. This version already no longer uses rgdal
. We hope make these changes with minimal breakage in backward compatibility by maintaining the conversion to sp
/raster
for functions from v < 4.0.0, but some backward incompatibilities will necessarily arise. In particular, LAS
will no longer inherit the sp::Spatial
class and will no longer contain a sp::CRS
but a sf::crs
and LAScatalog
will no longer be sp::SpatialPolygonDataFrame
. Our plan is (hopefully) to rebase lidR
in such a way that nobody will notice the changes expect users who dig a little deeper into the objects.
hexbin_metrics()
was an unused function and has been removed from lidR
. It can be retrieved in lidRplugins
Functions using the former namespace such as lassomething()
that were renamed into verb_noun()
in version 3.0.0 now throw a warning. In v3.0.0 they were still usable for backward compatibility but not documented. In v3.1.0 they printed a message saying to move on to the new namespace. Now in 3.2.0 they throw a formal warning saying to move on to the new namespace. They will throw an error in the next version.
classify_poi()
. New function capable of attributing a class of choice to any points that meet a logical criterion (e.g. Z > 2) and/or a spatial criterion (e.g. inside a polygon). For example, the following will attribute the class "high vegetation" to each non-ground point that is not in the lake polygon.
las <- classify_poi(las, LASHIGHVEGETATION, poi = ~Classification != 2, roi = lakes, inverse = TRUE)
LAScatalog
rbind()
for LAScatalog
.projection()<-
and crs()<-
for LAScatalog
. Those two functions were already working in previous versions but in absence of dedicated functions in lidR the functions that were actually called were raster::projection()
and raster::crs()
thanks to class inheritance. However the functions from raster
do not support crs
from sf
or numbers as input. Adding a dedicated function in lidR brings consistency between LAS
and LAScatalog
(#405):
projection(ctg) <- st_crs(3625)
# or
projection(ctg) <- 3625
ctg@chunk_options$drop
. This generates regions that won't be processed. This option accepts a vector of chunk IDs that are dropped and is thus versatile, but its main role is to allow restarting a computation that failed. We consequently introduced the function opt_restart()
. Let's assume that the computation failed after few hours at 80% in chunk number 800. Users get a partial output for the first 799 chunks but chunk 800 has a problem that can be solved. It is now possible to restart at 800 and get the second part of the output without restarting from 0:
output <- catlog_apply(ctg, myfun, param)
# Failed after 80%, 'output' contains a partial output
# Fix the trouble
opt_restart(ctg) <- 800
output2 <- catlog_apply(ctg, myfun, param)
# Merge 'output' and 'output2'
LAScatalog engine
and the manual LAScatalog-class
were updated to reflect these featuresLASheader
LASheader()
can now create a LASheader
object from a data.frame
. This addition aims to facilitate the creation of valid LAS
objects from external data.las_check()
can now check a standalone LASheader
las_check(las@header)
LAS
LAS
now automatically fixes the font case of attributes names to match the naming convention of the rlas
package. This simplifies the creation of compatible objects from non-LAS file sources.
data <- data.frame(x = runif(10), Y = runif(10), z = runif(10), pointsourceid = 1:10)
las <- LAS(data)
#> Attribute 'x' renamed 'X' to match with default attribute names.
#> Attribute 'z' renamed 'Z' to match with default attribute names.
#> Attribute 'pointsourceid' renamed 'PointSourceID' to match with default attribute names.
las$PointSourceID
#> [1] 1 2 3 4 5 6 7 8 9 10
Full waveform: with most recent versions of the rlas
package, full waveform (FWF) can be read and lidR
provides some compatible functions. However the support of FWF is still a work in progress in the rlas
package. How it is read, interpreted and represented in R may change. Consequently, tools provided by lidR
may also change until the support of FWF becomes mature and stable in rlas
.
interpret_waveform()
to transform waveform into a regular point cloudW
for parameter select
in readLAS()
Amplitude
in plot(las, color = "Amplitude")
that aims to be used with FWF.catalog_intersect()
now supports sf
, sfc
, Extent
and bbox
objects
Concave hull: lidR now includes its own C++ code to compute concave hulls using concaveman-cpp.
concaveman()
to compute concave hullsdelineate_crowns()
using concave hulls is now between 10 to 50 times faster.
LASfile <- system.file("extdata", "MixedConifer.laz", package="lidR")
las = readLAS(LASfile, select = "xyz0")
concave_hulls <- delineate_crowns(las, "concave")
# Before v3.2.0: 7.1 seconds
# From v3.2.0 : 0.2 seconds
grid_terrain()
with is_concave = TRUE
should also be faster.New function catalog_boundary()
to compute the actual shape of the point-cloud
In find_trees()
and segment_trees()
the bitmerge
strategy to generate robust unique IDs was not actually a valid and robust procedure. It had the advantage of generating integers but was not 100% unique. The probability to generate duplicates was low but we changed the strategy to use a true bit-merging procedure anyway. The new IDs thus generated are weird decimal number such as 5.001120e-310 but are guaranteed to be unique. The documentation has been updated to explain the method.
New algorithm random_per_voxel()
for decimate_points
that keep n points per voxel (#406).
3D rendering:
plot()
gains a new parameter voxels = TRUE
or voxels = 0.5
to render a point cloud with voxels. This is useful to render the output of voxelize_points()
or voxel_metrics()
, for example. This is computationally demanding and takes time so it should be reserved to small scenes with 30,000 or 40,000 voxels maximum, but note that there is no hard coded limit.
vm <- voxel_metrics(las, ~list(N = length(Z)), 8)
plot(vm, color = "V1", voxels = T)
plot()
.New function plot_metrics()
that wraps several other functions into one seamless function that extracts ground inventory plots, computes metrics for each plot and returns a ready to use data.frame
for statistical modelling.
New function point_eigenvalue()
that is equivalent to point_metrics(las, .stdshapemetrics)
but specialized, optimized and parallelized to be 10 times faster.
grid_metrics()
gains a new parameters by_echo
allowing users to compute the metrics for different types of echos independently. It is now possible to map e.g. mean(Intensity)
for first returns only + multiple return only + single return only. All metrics are computed in a single run and returned in a raster stack.
merge_spatial()
supports sfc
grid_density()
is 10 times fasterquantize()
now preserves NaN
values instead of converting them into minus infinity (#460).stdmetrics_i()
now fails with an informative message when the sum of intensities is greater than .Machine$integer.max
and becomes double
(#463)find_localmaxima()
respects the filter
argument. It was previously not considered.crayon
and hexbin
dependenciesRCSF
and rgeos
are now only suggested and they are consequently no longer installed by default with lidRrgdal
will be retired in 2024. Code using rgdal
internally has been removed. In many cases this will not change anything for users but in some cases it may fail when assigning an EPSG code to the LAS file. Also, old versions of rgdal
built with old versions of gdal
and proj
are no longer supported (#466)manual()
now uses the middle button to perform the selection. Historically the button was "right" but later the right button was added in lidR and attributed to the dragging action. By using "right" in this function this disabled the possibility to drag the scene. Consequently we changed the default to use the middle button. (#442).manual()
now removes all apices in the selection rectangle when removing some false positive (#445).catalog_apply
man pagevoxel_metrics()
with all_voxels = TRUE
did not work as expected. The insertion of empty voxels corrupted some of the real voxels. This bug lead to invalid output and some floating points precision errors lead to supernumerary voxels (#437, #439).grid_terrain()
used with a LAScatalog
no longer propagated the options. For example when using use_class = c(2L, 8L, 9L, 10L)
this was not propagated and the option was actually the default one i.e. use_class = c(2L, 9L)
. This bug was introduced in 3.1.0delineate_crowns()
now returns NULL
if the input point-cloud has only points with treeID = NA. It also triggers a warning. (#438).manual()
the function that allow for finding the trees manually was no longer working probably because of some slight modifications in the rgl
package.plot
function used to display the output of voxel_metrics()
now internally uses the same function than LAS
objects. This enhances the rendering using the clear_artifact
option by default and allows for a lot more flexibility in the rendering.button
in manual()
to choose which button to use.segment_trees()
now print a message if all points are NA
to suggest to use other parameterslas_check(..., deep = TRUE)
was not working in parallel (#411).las_check()
(#414)merge_spatial()
did not work with sf
objects.las_check()
introduces a new type of message called "message". Some message previously classified as "warning" are now classified as "message". Warnings are now displayed in orange and messages in yellow. The output of las_check()
has now 3 items instead of 2.stdmetrics_z
gains a new parameter zmin = 0
to control the lower bound of the integration for metrics zpcumx
(#424).max_cr_factor
in silva2019()
is now allowed to be in [0, inf[ instead of [0,1] (#417)sp
printing proj_create: crs not found
for non recognized EPSG codes and avoid throwing warning Discarded datum [...] in Proj4 definition
readLAScatalog()
throws a more informative error when attempting to read an non-existing folder.readXXXLAS()
now throws an error for LAScluster
(#430).stdmetrics
.LazyData
in DESCRIPTION
LASheader
has a new slot @EVLR
for the extended variable length records. print()
has been extended to display EVLR. While this change is compatible with rlas <= 1.3.9
it is only used with version of rlas >= 1.4.0
.lowest()
for decimate_points()
catalog_apply()
now works with cluster plan plan(cluster)
meaning that it can be used on HPC e.g. with MDPI. We took advantage of this bug to better detect the parallel strategy used and disable or not OpenMP. When lidR
is not able to figure out if the strategy involves multiple machines or multiple cores of a single machine, then a warning is thrown and OpenMP is disabled by security.
The parallel evaluation strategy was no recognized and lidR does not know if OpenMP should be disabled.
OpenMP has been disabled by security.
Use options(lidR.check.nested.parallelism = FALSE) and set_lidr_threads() for a fine control of parallelism.
spTransform()
have for consequences to make the function failing with error: Non quantizable value outside the range of representable values of type 'int'
.projection()
when using an epsg code as input (projection(las) <- 12345
).readLAScatalog()
now reads the WKT CRS of LAS files format 1.4. To support both EPSG and WKT the table of attribute of a LAScatalog
now has a column named CRS
that replace former column EPSG
.print()
for a LAScatalog
now prints the CRS exactly like print
for LAS
.options(lidR.check.nested.parallelism = FALSE)
was missing. Information can now be found in ?lidR-package
and ?lidR-parallelism
catalog_apply()
if lidR.check.nested.parallelism = FALSE
it now respects the input of set_lidr_thread()
instead of the output of get_lidr_threads()
. For example if set_lidr_thread(0)
it now propagates the information 0 (all cores) instead of the output of get_lidr_thread()
which might be e.g. 4 on the master worker but might be different on the slave workers. Similarly set_lidr_thread(20)
will request 20 cores to the workers even if get_lidr_thread()
returns 4 on the local machine.set_lidr_thread()
accepts inputs < 1 such as 0.5 or 0.25 to mean 'half' or 'quarter' of available cores.grid_density()
now returns 0 for pixels with 0 points instead of NA
which make more sense and corresponds to what should be expected.