This document evaluates algorithms for detecting individual trees from LiDAR data. It compares local maxima (LM) algorithms on raster and point cloud data to a new pattern recognition algorithm based on geomorphons. All algorithms were tested on a study area in Italy containing different forest structures. The pattern recognition algorithm detected trees most accurately but the point cloud LM algorithm performed best overall. Particle swarm optimization calibration improved detection rates over manual calibration. The algorithms show potential for estimating forest parameters like volume from remote sensing data at large scales.
Application of a Pattern Recognition Algorithm for Single Tree Detection from LiDAR Data
1. APPLICATION OF A PATTERN
RECOGNITION ALGORITHM FOR
SINGLE TREE DETECTION FROM
LiDAR DATA
FOSS4G-EU 2017 â 2017/07/21
A. Antonello, S. Franceschi,
V. Floreancig, F. Comiti(*), G. Tonon(*)
(*) Free University of Bolzano
2. INDEX
1. Research objective
2. Tools to extract vegetation from LiDAR data
3. Application on the study area
4. Conclusions
3. RESEARCH OBJECTIVE
Evaluate the use of LiDAR data to extract forest
parameters:
â
position and height of single trees
â
forest volume.
4. Extrapolation of the whole forest biometric information
(e.g. tree height, DBH, forest biomass) from remote
sensing data can be obtained through two approaches:
â
area-based approaches (AB): forest attributes are
estimated relating plot information to ALS data by
statistically procedure
â
individual tree crown (ITC) approaches: are aimed to
detect position and main characteristics of each single
tree. Single-tree records can then be aggregated at plot,
forest, watershed or regional scale.
EXTRACT VEGETATION FROM LiDAR
5. Extrapolation of the whole forest biometric information
(e.g. tree height, DBH, forest biomass) from remote
sensing data can be obtained through two approaches:
â
area-based approaches (AB): forest attributes are
estimated relating plot information to ALS data by
statistically procedures
â
individual tree crown (ITC) approaches: are aimed to
detect position and main characteristics of each single
tree. Single-tree records can then be aggregated at plot,
forest, watershed or regional scale.
EXTRACT VEGETATION FROM LiDAR
6. EXTRACT VEGETATION FROM LiDAR
â
three models: Local Maxima
raster and point cloud + pattern
recognition raster
7. â
three models: Local Maxima
raster and point cloud + pattern
recognition raster
â
matching procedure between
extracted and measured trees is
based on relative position
(distance) and height
EXTRACT VEGETATION FROM LiDAR
8. â
three models: Local Maxima
raster and point cloud + pattern
recognition raster
â
matching procedure between
extracted and measured trees is
based on relative position
(distance) and height
â
automatic calibration is based
on Particle Swarming Optimizer
(PS)
EXTRACT VEGETATION FROM LiDAR
9. â
three models: Local Maxima
raster and point cloud + pattern
recognition raster
â
matching procedure between
extracted and measured trees is
based on relative position
(distance) and height
â
automatic calibration is based
on Particle Swarming Optimizer
(PS)
â
three fitness functions based on
the statistical concept of
Information Retrieval: minimize
FP, minimize FN and the
harmonic mean of FP and FN
EXTRACT VEGETATION FROM LiDAR
10. â
Local Maxima algorithms identify points with the highest
elevation value within a given moving window
RASTER POINT CLOUD
Moving window circular circular
Center of window each raster cell all point of the cloud
Radius constant constant
Data used DSM point cloud + DTM
Calibration params â
maximum radius
â
height difference for
branches
â
min tree height
â
maximum radius
â
height difference for
branches
EXTRACT VEGETATION FROM LiDAR
11. â
Pattern Recognition model is based on the geomorphon
algorithm (Jasiewicz, 2013)
â
uses a pattern based classification of landforms: characterization
of the local surface using the line-of-sight principle (LOS)
â
LOS describes the unimpeded view or access from one point to
another point across a terrain or surface
â
LOS is used normally to understand the visible and obstructed
(non visible) points in terrain analysis
â
498 possible different patterns of which 10 are the mostly
common â 4 investigated for tree detection
EXTRACT VEGETATION FROM LiDAR
12. EXTRACT VEGETATION FROM LiDAR
Parameters of the Geomorphon model are:
â
maximum lookup distance: in the 8 main directions
â
flatness threshold: vertical angle (angle between the horizontal
plane and the line connecting the central cell with the point
located on the profile at the line-of-sight distance or at the
maximum allowed distance)
13. â
additional checks have been implemented for filtering peaks
which are surrounded by:
â
no more than 2 not valid cells (otherwise it is considered a point
on the boundary of a crown)
â
no other peaks
â
not a coexistence of pits, valleys, spurs and hollows
â
treetops extracted from the geomorphon peaks do not have in
their surrounding other sharp morphologies, there should be
slopes, ridges or shoulders which indicate a continuity of the
shape that generates the central peak.
EXTRACT VEGETATION FROM LiDAR
14. EXTRACT VEGETATION FROM LiDAR
â
raster based using as reference surface the CHM
â
calibrated parameters:
âą
the maximum lookup distance for the LOS
âą
the flatness threshold (difference between the nadir and the
zenith angles) for the LOS
âą
the height difference for filtering branches (filter on breaching
patterns of trees)
â
output is a set of treetops position with the height as attribute
15. FALSE POSITIVE
extracted by the software
but not existing in reality
mapped trees
extracted trees
FALSE NEGATIVE
existing but not extracted
by the software
TRUE POSITIVE
extracted by the software
and matched with measured
â
AGB is calculated using allometric functions and hypsometric
relationship where the parameters are extracted from field data
V=
aâ Ïâ D
2
4
â H
DLIDAR=bâ H
2
+c ËH +d
STUDY AREA a b c d
Val Aurina 0.0000368048 0.0096 1.298 0.00
EXTRACT VEGETATION FROM LiDAR
16. STUDY AREA
AURINA VALLEY high local variety in forest structure
â
Norway spruce (Picea abies)
â
Larch (Larix decidua)
â
Stone pine (Pinus cembra)
AREA = 10 km2
LiDAR SURVEY OF 09/2012
points density = 10 p/m2
FIELD SURVEY SUMMER 2013
3 FOREST STRUCTURES:
monoplane â biplane â multilayer
12 circular plots
radius = 15 m
(position + height +
DBH + species)
17. CHARACTERISTIC Monoplane Biplane Multilayer
Tree density (n/ha) 427.2 473.9 331.0
Basal area (m2
/ha) 26.7 39.9 34.6
Average Basal area (m2
) 0.064 0.101 0.114
Volume AGB (m3
/ha) 121.0 387.4 283.7
STUDY AREA
18. RESULTS
Simulation have been done using the Particle Swarming automatic
calibration. The classification of results considered:
â
3 different models
âą
LM raster based
âą
LM point cloud based
âą
Geomorphon raster based
â
3 different fitness functions
âą
minFP
âą
minFN
âą
harmonic mean of FP and FN
â
3 different forest strutures
âą
monoplane
âą
biplane
âą
multilayer
19. Mean detection rates of treetops
Rmat =
Ntp
Nmap
Rcom=
Nfp
Nmap
Rext=
Next
Nmap
RESULTS
METHOD Rext Rmat Rcom Rom
PS_rast 0.95 0.74 0.21 0.26
PS_pc 0.88 0.77 0.11 0.23
PS_geo 1.01 0.80 0.22 0.20
AVERAGE 0.95 0.77 0.18 0.23
22. JGRASSTOOLS LIBRARY
â
an open source geospatial library focused on hydro-
geomorphological analysis and environmental modeling
â
development started in 2002 at the University of Trento,
Department of Civil and Environmental Engineering
â
in 2012 started the development of LESTO (LiDAR
Empowered Scientific Toolbox Open Source) in
collaboration with the University of Bolzano
â
from 2015 integrated as Spatial Toolbox in gvSIG
â
available for installation through the gvSIG Update
Manager as a plugin
23. JGRASSTOOLS LIBRARY
Models are grouped in sections
and subsections. Main sections
are:
â
HortonMachine:
geomorphology analysis
â
Raster and vector processing
â
Mobile tools: support for
Geopaparazzi application for
digital field mapping
â
LESTO: LiDAR analysis
(forestry)
25. CONCLUSIONS
â
the use of PSO iterative parametrization compared
with commonly used LM algorithms with manual
calibration performs better in terms of detection rates
â
the new model based on LM for point cloud (PC) is the
one that performs better in all the analyzed conditions
and also for different trees species (conifers and
broadleaves)
â
the new model based on pattern recognition
(geomorphon) is a valid alternative to standard LM if
only raster based input data are available
â
the fitness function based on the harmonic mean of
FP and FN seems to be the most accurate for both
volume and number of trees
26. Andrea Antonello, Silvia Franceschi
V. Floreancig, F. Comiti, G. Tonon
THANKS FOR THE ATTENTION!