Single tree reconstruction - foto.aalto.fi · VRS Virtual Reference Station WGS World Geodetic...
Transcript of Single tree reconstruction - foto.aalto.fi · VRS Virtual Reference Station WGS World Geodetic...
SINGLE TREE FEATURE EXTRACTION FROM AIRBORNE LASER SCANNER DATA
APPLYING VECTOR MODELS
Thesis for degree of Licentiate of Science In Technology Submitted to the Department of Surveying at Helsinki University of Technology Helsinki, August 2006 Ulla Pyysalo
Supervisor: Prof., Dr. Henrik Haggrén Instructor: Docent, Dr. Hannu Hyyppä
HELSINKI UNIVERSITY OF TECHNOLOGY ABSTRACT OF THE LICENTIATE’S THESIS
Author: Name of the thesis: Date: 28.8.2006
Ulla Pyysalo Single Tree Feature Extraction from Airborne Laser Scanner Data applying Vector Models Number of pages: 78
Faculty: Department of Surveying
Professorship: Photogrammetry and Remote Sensing
Supervisor: Prof. Dr. Henrik Haggrén Instructor: D.Sc. Hannu Hyyppä
The objective of the study was to develop 3D vector models of single trees from laser scanner data in order to derive geometry features. The vector model production was implemented in the following stages: point classification, digital terrain model production, extraction of points from each tree, and vector model creation. The extracted features were tree height, crown height, trunk location, and crown profile. The method was applied at the Otaniemi test site. Reference materials from trees in the test area were acquired in field measurements by tacheometer and hypsometer. In order to derive crown shape information, a new method based on side view images of trees was tested. Trees were imaged from two directions with determined geometry and rectified to a plane parallel with the tree trunk, from which crown shape parameters were acquired. The features extracted from laser models were compared to the reference material. The results suggested that tree shape is underestimated in laser derived models in both vertical and horizontal direction. The results showed that tree location could be extracted with an accuracy of 2 m and tree heights with an accuracy of 1.5 m. Manual extraction of tree points resulted in better accuracy than segmentation-based extraction. The crown shape was sufficiently accurate at the top of the trees, but accuracy decreased significantly towards the tree base. The impact of laser scanning parameters on vector model production was considered by analysing error sources in different data sets and applying data simulations. The point density and scanning pattern were found to have the most profound effect on the segmentation and crown shape accuracy. However, the vertical feature, tree height, was obtained from all data sets to an almost similar accuracy regardless of point density.
Keywords: Laser scanning, Vector models, Side view images, 3D features
Language: English
2
TEKNILLINEN KORKEAKOULU LISENSIAATTITYÖN TIIVISTELMÄ
Tekijä: Työn nimi: Päivämäärä: 28.8.2006
Ulla Pyysalo Single Tree Feature Extraction from Airborne Laser Scanner Data applying Vector Models Sivumäärä: 78
Osasto: Maanmittaus Professuuri: Fotogrammetria ja kaukokartoitus
Supervisor: Prof. Dr. Henrik Haggrén Instructor: D.Sc. Hannu Hyyppä
Lisensiaattityön tavoitteena oli kehittää menetelmä yksittäisten puiden vektorimallien muodostamiseksi piirteiden irrotusta varten. Vektorimallin muodostus tapahtui työvaiheissa, jotka olivat pisteiden luokitus, alueen korkeusmallin luonti, yhdestä puusta heijastuneiden pisteiden irrotus ympäristöstä ja vektorimallin muodostus. Malleilta irroitettiin puun geometrisiä piirteitä, jotka olivat puun ja latvuksen pituus, rungon sijainti ja latvuksen muoto. Menetelmää kokeiltiin Otaniemen koealueen puihin viidellä eri laserkeilainaineistolla. Puista kerättiin referenssiaineistoa takymetri– ja hypsometrimittauksin. Puiden latvuksen muodon määrittämiseksi kehitettiin sivuvalokuvaukseen perustuva menetelmä, jossa kuvat otetaan määrätyllä geometrialla ja oikaistaan rungon suuntaiselle tasolle mittausta varten. Vektorimalleilta irrotettujen piirteiden tarkkuutta arvioitiin vertaamalla niitä referenssiaineistoon. Tutkimuksessa selvisi, että puun muoto aliarvioidaan lasermalleilla sekä pysty- että vaakasuunnassa. Puun rungon sijainti määritettiin malleilta n. 2 m tarkkuudella ja korkeus n. 1.5 m tarkkuudella. Puun latvuksen yläosan muoto voitiin kuvata malleilla tarkasti, mutta tarkkuus huononi siirryttäessä alaspäin kohti puun juurta. Laserkeilauksen parametrien vaikutusta vektorimallien muodostukseen arvioitiin tarkastelemalla virhelähteitä eri parametreillä mitatuissa aineistoissa sekä aineiston simulaatioilla. Pistetiheydellä ja keilauskuviolla huomattiin olevan suurin vaikutus segmentointitarkkuuteen ja latvan muodon määritykseen. Puun pituus taas määritettiin eri aineistoista muodostetuilta malleilta lähes samalla tarkkuudella pistetiheydestä riippumatta.
Avainsanat: Laserkeilaus, vektorimallit, yksittäiset puut, sivukuvaus, 3D piirteet
Työn kieli: Englanti
3
PREFACE
The thesis is result of work, which has been conducted in the Institute of Photogrammetry and
Remote Sensing Helsinki University of Technology between years 2000-2006. I would like to
thank my supervisor professor Henrik Haggrén for opportunity to work in the laboratory these
years and encouragement for the research scientist carrier. My instructor D.Sc. Hannu Hyyppä
I would like express my special thanks for great support, inspiring discussions and positive
energy during the research period, in addition to the trumpet-shaped chanterelle he delivered
me in autumns.
I would also like to thank professor Matti Maltamo and D.Sc. Jussi Heikkinen for helpful
comments and reviewing the thesis. Furthermore I am grateful to professor Juha Hyyppä for
helpful comments and knowledge concerning laser scanning, Arttu Soininen for possibility to
use TerraScan- software and Jaakko Järvinen and the team for carrying out field
measurements. The colleagues and friends in the Institute of Photogrammetry and Remote
Sensing I thank for the pleasant and comfortable working environment. Especially I thank my
office roommate Lic.Sc. Petteri Pöntinen for seven enjoyable years we shared the room.
During the years I received financial support from the Ministry of Agriculture and Forestry,
the Finnish Academy, the Finnish Cultural Foundation, Jenny and Antti Wihuri Foundation,
and the Nature Resource Foundation. I thank for it.
Finally, I would like to thank family and friends for support and encouragement.
4
CONTENTS LIST OF ABBREVATIONS 7
LIST OF SYMBOLS 8
1 INTRODUCTION 9
1.1 Background and motivation 9
1.2 Previous studies 11
1.3 The objectives and structure of the thesis 12
2 AIRBORNE LASER SCANNER 15
2.1 Overview of laser scanning measurement 15
2.2 Laser range finder 16
2.3 Scanner 18
2.4 Positioning unit and data controlling and possessing 19
2.5 Additional components 19
2.6 Laser scanning accuracy 20
2.7 Point density in laser scanner data 21
3 MATERIALS 23
3.1 The Otaniemi test site 23
3.2 Laser scanner implementations 23
3.3 Laser scanner data sets 25
3.4 Acquisition of reference material 26
4 SIDE VIEW IMAGES 28
4.1 Motivation 28
4.2 Capturing side view images at the Otaniemi test site 29
4.3 Image preprocessing 30
4.4 Image measurements 31
4.5 Determining the trunk tilt 32
4.6 Error analysis 34
4.7 Method analysis 37
5
5 THE VECTOR MODEL PRODUCTION METHOD 38
5.1 Preprocessing of laser data 38
5.2 Point classification and DTM 39
5.3 Tree crown delineation and point extraction 41
5.4 Vector model processing method 43
5.5 Tree model feature extraction 46
6 RESULTS AND THEIR ANALYSIS 48
6.1 The evaluation of DTM accuracy 49
6.2 Single tree point extraction by tree delineation 50
6.3 Tree heights and crown heights 53
6.4 Locations of the tree trunks 56
6.5 Crown width analysis 59
7 THE IMPACT OF LASER SCANNING PARAMETERS ON THE APPLICATION
OF VECTOR MODELS 64
7.1 Point density and reconstruction 64
7.2 Effect of overlapping flight lines 66
7.3 Beam sizes and biases 67
8 SUMMARY AND CONCLUSIONS 68
8.1 Summary and conclusions 68
8.2 Discussion 70
8.3 Future work 71
REFERENCES 73
6
LIST OF ABBREVATIONS
3D Three-Dimensional
ALS Airborne Laser Scanner
CHM Canopy Height Model
CW Continuous Wave
DGPS Differential Global Positioning System
DSM Digital Surface Model
DTHM Digital Tree Height Model
DTM Digital Terrain Model
GIS Geoinformation System
GPS Global Positioning System
IFOW Instantaneous Field of View
IMU Inertial Measurement Unit
KKJ Finnish National Coordinate System
LASER Light Amplification by Stimulated Emission of Radiation
PRF Pulse Repetition Frequency
STD Standard deviation
TIN Triangulated Irregular Network
VRS Virtual Reference Station
WGS World Geodetic System
7
LIST OF SYMBOLS R range
Lt pulse time of flight
pφ phase difference
n number of full wavelengths
T travelling time of full wavelength
c speed of light 300 000 000 m/s
ggg zyx ,, ground co-ordinate system axes
ϖ rotation around -axis gx
φ rotation around -axis gy
κ rotation around -axis gz
x, y image co-ordinates
X, Y co-ordinates on rectification plane
21210210 ,,,,,,, ggfffeee unknown parameters of projective transformation
s scale
h point elevation from the ground
tt yx , trunk co-ordinates
r range between trunk and point
α azimuth
v trunk vector
sX1 distance between trunk and tree top in horizontal direction in image 1
sX2 distance between trunk and tree top in horizontal direction in image 2
),( 11 cc yx image plane and polygon intersection point 1
),( 22 cc yx image plane and polygon intersection point 2
dw1 distance between trunk and intersection point 1
dw2 distance between trunk and intersection point 2
8
1 INTRODUCTION
1.1 Background and motivation
The rapid development of laser scanning has offered new challenges in the characterization of
the geometry of individual trees. When a forest is measured by airborne laser scanning, the
tree canopy reflects the transmitted laser pulses. Several echoes can result from one
transmitted laser pulse, since laser beams can illuminate vegetation at various distances. If the
measurement point density is as high as ten points per square meter, vegetation objects, such
as trees and bushes, can be distinguished visually from the measurement data on the basis of
their geometric form (see Figure 1.1).
Figure 1.1. Side view of the laser scanner data of a forested area.
The laser scanner point clouds collected from the air enable new approaches to forest
inventory. Conventional forest inventory methods can be roughly divided into two categories:
region-based, such as stand- and plotwise inventory, and treewise inventory, which is based
on individual trees. Traditionally airborne materials, aerial photographs or satellite images,
have been extensively applied to region-based inventory, since the advantage of these data is
the recording of large areas in a single shot. However, the disadvantage of satellite imagery is
its small scale, i.e., several individual trees mixed into one pixel, and its limited quality in
recording three-dimensional (3D) forest geometry. Today, in order to collect attributes related
to individual trees, field measurements have to be conducted. Standwise forest attributes can
be calculated on the basis of individual trees and sample plots. This approach, however, is
time-consuming and expensive.
9
The airborne laser scanning of forest enables inventories of large areas while simultaneously
providing three-dimensional (3D) information about individual trees and their parameters,
such as tree height and crown width. If needed, the results from individual trees may be
generalized to stand level by applying a stand boundary map. Increasing the flying altitude
increases the width of the measurement stripe, which can be measured with single fly pass,
thus the energy of light restricts high altitude measurement. However, the disadvantage is that
the airborne view of the forest is limited, since tree crowns shadow the lower parts of the tree
and trunk (Figure 1.2). A tree is a vertically oriented object (Figure 1.3) and the majority of its
3D surface is unseen when viewed at a steep angle. In order to provide measurements from
the shadowed area, the laser light should be capable of penetrating through the canopy and
also providing energy to carry the reflected pulse back to the receiver. In reality, it is
questionable whether 3D tree geometry from the lower parts of the tree crowns is recorded by
airborne laser. In addition, laser pulses are not directed to treetops or other targets, but
distributed to the measurement stripe by means of a scanning mechanism. Therefore all tree
parameters are retrieved from laser-derived models or surfaces as secondary parameters
instead of as direct measurements. This thesis focuses on answering which individual tree
parameters may be extracted from the data, and how accurate values of parameters can be
assumed to be.
Figure 1.2. Airborne view. Figure 1.3. Terrestrial view (Rönnholm, 2004).
The approach, laser scanning assisted forest inventory, provides information that has not been
obtained previously. Tree geometry parameters below the canopy surface are problematic
even when traditional terrestrial measurements are applied. Trunk locations may be
10
determined by tacheometer measurements and tree heights may be obtained by hypsometer
measurements. The crown widths and profiles, however, are not easily registered with a
tacheometer or other positioning instrument. In order to gain this same information in
terrestrial measurements, a new method, side view imaging of the trees, was tested in the
thesis. The method also provides reference material for laser-derived tree attribute
comparisons.
1.2 Previous studies
The promising nature of airborne laser scanner data has inspired several research scientists to
explore the applicability of laser material to forest inventory. The studies started around 1980
(Solodukhin et al., 1977; Nelson et al., 1984; Aldred and Bonnor, 1985; Maclean and Krabill,
1986) and concentrated on using a profiling system for forest height, stand density, tree
species, and biomass estimation. The basics of using laser measurements for forest inventory
were established at that time.
There are numerous studies concentrating on certain individual tree attributes. The most
researched parameter is tree height, which may be applied as an input parameter in timber
volume estimation models. Hyyppä and Inkinen (1999) were the first to demonstrate the
possibility of measuring treewise information using airborne laser scanning and of adopting
the retrieved parameters (height, crown width) in forest inventory calculations. Their results
indicated that the height of trees may be measured to an accuracy of better than 1 m even
though tree heights are underestimated. Similar results have been obtained later by Persson et
al. 2002, Gaveau and Hill, 2003, Leckie et al., 2003, Yu et al., 2004, Maltamo et al., 2004a. In
these studies, the typical approach has been to process the forest canopy as a surface that is
obtained from laser scanner data. The forest canopy surface is constructed as a triangulated
irregular network (TIN), contour or grid model. Magnussen and Bouldewyn (1998)
introduced a geometrical model that successfully predicted the mean difference between the
laser canopy heights and the mean tree height. The model explained why estimation of stand
heights from laser scanner data based on maximum canopy height value in each cell of a fixed
area grid (e.g. Næsset 1997b) has been successful in practice. Magnussen, Eggermont and
LaRiccia (1999) introduced two recovery models that could be used to obtain tree heights
from laser height measurements.
11
In Yu et al. (2004), the capability to measure forest height growth and to detect harvested
trees from multitemporal laser surveys was demonstrated.
There are few studies focusing on tree geometry below the canopy surface. The vertical
structure of forest has been studied by Friedlaender and Koch (2000) by analysing statistically
the penetration of a laser pulse in forest canopy. In Maltamo et al. (2004), the existence and
number of suppressed trees was examined. This was carried out by analysing the height
distributions of reflected laser pulses. Height percentiles of the distribution of canopy heights
have been also used as predictors in regression models for the estimation of mean tree height,
basal area and volume (e.g. Lefsky et al., 1999; Magnussen et al., 1999; Means et al., 2000;
Naesset, 1997a,b; Naesset and Okland, 2002; Naesset, 2002). In these approaches the results
are gained for forest stands instead of individuals trees. In Pyysalo and Hyyppä (2002),
another approach was introduced to describe the tree crown using vector polygons bounding
crown sides and line estimating the shape and location of the trunk. Finding tree locations can
be also obtained by detecting image local maxima using image processing methods (e.g.
Geogeon and Moore 1989). These so-called segmentation-based methods are discussed in
more detail in chapter 5.3.
1.3 The objectives and structure of the thesis
The objective of this study was to extract features of individual trees from laser scanner data
applying vector models. Vector models were implemented in the following stages: point
classification, digital terrain model production, extraction of points from each individual tree,
and vector model creation. Different laser scanned data sets were applied in this study, which
enabled the comparison of data sets and their parameters. The sub-objectives of this thesis
were:
1. to develop a method for constructing vector models of single trees from airborne laser
scanner data
2. to develop a method for measuring reference data from side view images
3. to evaluate the accuracy of the tree geometry features derived from the vector model
and
4. to consider the effect of different laser scanning parameters in the vector model
construction process.
12
The input data, workflow steps and sub-objectives are illustrated in Figure 1.4.
The topic of digital terrain model (DTM) production is discussed briefly in this thesis, since
errors in ground altitude have an impact on tree height extraction. Previous studies concerning
DTM are introduced in chapter 5.2.
Laser Sacnnerdata
Side view images
Hypsometermeasurements
Tacheometermeasurements
MAT
ERIA
LSM
ETH
OD
S
Objective 1:Applying
vector models
Objective 3:Comparison of features extracted from vector models to reference materials
Objective 4:Analysis of the impact of laser scanning parameters
Objective 2:Applying
side view image measurements
Figure 1.4. Objectives of the thesis.
However, this thesis is limited to the discussion of features that are directly measurable from
laser scanner data, and therefore tree parameters such as diameter at breast height or volume,
which are typically estimated from other parameters, are not considered.
This thesis has eight chapters and a structure with introduction, materials and methods
followed by results. The introduction in Chapter 1 includes a short description of the work,
objectives of the thesis, and a summary of previous studies. Chapter 2 presents the technical
13
description of laser scanner and system parameters. The test area, three different laser scanner
systems, and laser scanner data sets applied in the study are introduced in Chapter 3.
The first method to be explained is the side view imaging of trees for collecting reference data
in Chapter 4. The second method, vector model creation, is described in Chapter 5. These
chapters are followed by the results of the comparison of applied models to reference data and
the analysis of the results in Chapter 6. The laser scanner parameters that affect the
reconstruction are discussed in Chapter 7, and finally conclusions and discussion are provided
in Chapter 8.
14
2 AIRBORNE LASER SCANNING 2.1 Overview of laser scanner measurement An airborne laser scanner (ALS) is a helicopter- or airplane-mounted measurement system
that provides 3D point clouds from the measurement area. The laser transmitter emits a laser
pulse, typically at the specific wavelength, receives the backscattered return, and determines
the distance to the object based on the time of flight. With the scanning mechanism, the area
is covered with laser pulses in the across-track direction. The ALS system also includes a
positioning system, so that measured distances can be calculated to vector origins with co-
ordinates.
Scanning angle
Distance vector
Vector origin DGPSIMU
Beam divergence angle
Figure 2.1. Principle of laser scanning
The name LASER is an acronym of the words Light Amplification by Stimulated Emission of
Radiation. Laser light has physical characteristics suitable for remote sensing ranging, as the
signal is highly coherent and therefore powerful and directional. The predecessors of today’s
laser scanners include the LIDAR (profilometer) and bathymeter, for example (Rönnholm,
2005). The first airborne lasers were adjusted for direct measurement in one stable direction
only, and therefore the measurement pattern from one flight line was a profile. In airborne
3D-profilometer a laser-camera combination was introduced; the camera captured surface
illuminated by laser line (Haggrén et al., 1995). The combination of inertia measurement and
Global Positioning System (GPS) in 1980 increased the positioning accuracy, and a scanning
mechanism to distribute pulses across the track was applied. The first laser scanner came on
the market in 1990, and since then they have been used increasingly often for operational
surveying purposes. In recent years, the development of laser scanners has introduced new
15
measurement modes, increased measurement frequency, and included digital cameras in the
systems.
The laser scanner implementation usually includes four components: (1) laser range finder
(LRF) measurement unit, (2) positioning unit, (3) scanner and (4) data control and possessing
unit (Figure 2.2) (Wehr and Lohr, 1999). In addition, digital cameras are included in several
systems. This chapter introduces these components and the principle of laser measurement.
Laser scanning accuracy and point density are considered in sections 2.6 and 2.7.
DGPS
IMU LASER CAMERA
POSITIONING UNIT
DATA CONTROLLING AND POSSESSING UNIT X,Y,Z
Figure 2.2. Laser system components (Wehr & Lohr, 1999)
2.2 Laser range finder
The laser ranging unit contains the laser transmitting and receiving components. Most
airborne lasers are pulse lasers, which emit short pulses and measure the time of flight of each
pulse. A ranging vector length is derived according to Formula 2.1. The other principle of
laser measurement consists of transmitting and receiving a continuous sinusoidal modulated
signal. In such continuous wave (CW) systems, the time of flight is derived from the phase
difference between the transmitted and received signal and the number of full wavelengths
(Formula 2.2) (Wehr and Lohr, 1999). Due to the limited amount of CW lasers used for
airborne surveying purposes, the following text concentrates on pulse laser measurement
properties.
LtcR ⋅⋅=21 (2.1)
TnTt pL +⋅=
πφ2
(2.2)
16
R range, c speed of light 300 000 000 m/s,
time of flight of the laser pulse Lt
pφ phase difference,
n number of full wavelengths T travelling time of full wavelength The emitted pulse widens during the time of flight of the laser pulse according to the beam
divergence angle and illuminates an elliptical area (equivalent to an instantaneous field of
view (IFOW)) when reaching the surface (Figure 2.3). The beam divergence typically ranges
between 0.3 and 2 mrad depending on the system (Wehr & Lohr, 1999). In some laser
scanner implementations, beam divergence is an adjustable parameter. A large IFOW is
desirable when the objects to be scanned are small in size, for example tree tops and power
lines. In contrast, the positional accuracy of a laser pulse is inversely proportional to the
illumination area, and therefore widening the IFOW will decrease the accuracy.
Beamdivergence
angle
IFOV
Pulse rise time
time
Figure 2.3. IFOV Figure 2.4. Pulse rise time.
The receiver component has been mounted so that transmitted and received pulses share the
same optical path (Wehr and Lohr, 1999). The field of view (FOV) in reception must not be
smaller than that in transmission. However, the FOW value is limited by diffraction, which
causes image blurring (Wehr and Lohr, 1999). A pulse is registered as a return signal when
the threshold of amplitude is reached or according to some other algorithm (e.g. the slope of
the return signal). The reception unit registers one or several returns from the received echo
by analyzing the return signal shape. First, last and even some middle echoes are registered by
some systems. If the whole pulse is reflected from solid surfaces perpendicular to the
direction of pulse transmission, the return signal has one steep peak and only one point is
registered (Figure 2.4). However, even a very short emitting period, such as 4 ns, produces a
17
light line of 1.2 m, and the pulse is typically reflected from multiple surfaces at different
distances and the return signal shape is more complex (Figure 2.5).
Most systems provide intensity information of the return signal. These values are typically not
absolute but scaled to a number of interval that the data possessing unit is capable of storing.
During the last few years, a new recording mode has been introduced, i.e. full waveform.
Instead of registering just one or few intensity values from the echo, for example 128 samples
from the signal are collected (Figure 2.5). With recording interval for 0.5 ns the range
between adjacent samples becomes 15 cm and the length of whole measurement vector 19.2
m. Within this space, the interaction of pulse to reflecting objects is recorded in one
waveform measurement.
Ground, only echo
Vegetation, two registered echoes
Figure 2.5. Signals from two different targets, ground only and forest canopy and ground. The horizontal red and
green lines are locations where instruments determined returned echoes.
2.3 Scanner
In order to distribute pulses across the area to be covered, a scanning mechanism is applied.
The second dimension in the imaging is achieved when the platform moves forward. The
measurement frequency refers to the pulse repetition frequency (PRF), and the scanning
frequency indicates how many across-track scans are applied during one second. The most
common scanning patterns are parallel lines (i.e. pushbroom scanning), zig-zag pattern and
elliptical scans (i.e. conical scanning with fixed incidence angle) (Figure 2.6). The technical
implementation in zig-zag scanning is accomplished by shifting the direction of pulse
transmission with swivelling mirrors or prisms, and in pushbroom scanning by transmitting
pulses through linear array fibres (Wehr & Lohr, 1999). The scanning may be one- or two-
directional depending of the deflection unit. The scanning frequency and pulse distribution
18
function are results of the beam distribution mechanism. Point density across the scan is not
constant, but increases or decreases from nadir below the platform towards the cover area
(Figure 2.6).
Figure 2.6. Left: distorted conical scanning, middle: two-directional zig-zag –pattern, right: one-dimensional
zig-zag pattern.
2.4 Positioning unit and data controlling and possessing
The positioning system of a laser scanner implementation has two units, the differential global
positioning system (DGPS) and the inertia measurement system (IMU). In addition, a DGPS
instrument should be placed near the study area or a virtual GPS solution should be used.
These together provide orientation and vector origin for each measurement. Laser range
measurements are synchronized with positioning unit values according to a time stamp
attached to each data item. The frequency of DGPS measurement is typically 1 Hz. Therefore
DGPS location measurements are supported by higher-frequency IMU. Finally the co-
ordinates for each vector origin are interpolated from the closest known locations. During the
flight, collected material is stored in the fourth component of the laser system, the data
controlling and possessing unit. Two examples of computer configuration are found in
Baltsavias (2000) and Wehr and Lohr (1999).
2.5 Additional components
Medium-format digital cameras, where image frame is for example 2032×3056 pixels, can be
included on several laser scanning implementations. Aerial digital images are taken during the
laser range capture and synchronized with location information from the positioning unit and
laser measurements using a time stamp attached to each image. In order to solve orientation of
images, co-ordinates for image capture locations are determined. A typical image capture
interval is 2-3 seconds, depending on the flying altitude and velocity.
19
2.6 Laser scanning accuracy
The accuracy of a laser-derived 3D point depends on three main factors: (1) ranging accuracy,
(2) accuracy of the distance vector origin and (3) direction of the laser pulse (Baltsavias,
2000). In addition, preprocessing steps such as strip matching and co-ordinate transformation
influence the final co-ordinates.
The ranging accuracy is influenced by the pulse transmission and receiving distance and
cover material of the objects (Baltasavias, 2000). The optimal way of transmission is to send
steep and short pulses. The pulse rise time, which is independent of the pulse width,
determines the time required for an emitted beam to increase from 10 % to 90 % of the
maximum power value (Figure 2.4). Even a short rise time, such as 1 ns, will correspond to
15 cm in range. Reflections from objects shorter than the pulse length are overlapping, and
therefore the challenge is in analysing the returning signal shape in order to determine the
return location. The power of the return signal corresponds to the distance and incident angle,
but also to the reflectivity of the surface material. Glass surfaces, for example, may reflect
pulses out of the area of reception, and a forest canopy surface may cause several weak
signals, which are mixed with background noise. According to Wehr and Lohr (1999), the
ranging accuracy depends on signal-to-noise ratio (S/N), which is affected by factors such as
the power of the received signal, input bandwidth, background radiation, amplifier noise etc.
The positional accuracy of a single laser pulse is inversely proportional to the size of the
IFOV, since all reflections from the IFOV area are considered to originate from the pulse
transmitting direction (Figure 2.3). As mentioned earlier, the beam divergence angle is
adjustable in some systems. In other systems the value is constant, and therefore the size of
the beam ellipse may only be changed by changing the flying altitude.
The accuracy of the distance vector origin is a result of the positioning and orientation
measurements, i.e. the behaviour of DGPS and IMU. According to the POS unit
manufacturer, the accuracy of their combination should be better than 0.1 m. (Wehr and Lohr,
1999). This accuracy is achieved when the GPS measurement is supported by a DGPS
receiver stationed to a local ground control point or a Virtual Reference Station (VRS) is
applied from stable GPS stations. In addition, systematic parameters must be considered for
each surveying flight. These are the three mounting angles - roll, pitch and heading - and the
20
shift between the fixed platform co-ordinate system and the positioning unit and laser (Wehr
& Lohr, 1999). Parameter values are usually solved in preprocessing by applying calibration
data, which is typically collected by flying over flat areas, i.e., football fields, airports or
water surfaces, from different directions.
The determination of the laser pulse transmitting direction also depends on the quality of INS.
The quality is affected by the INS measurement frequency, postprocessing method, and
integration of GPS. (Baltsavias, 2000)
If all error sources are combined, the typical horizontal accuracy of single laser measurement
is approximately 20-30 cm and the vertical accuracy is approximately 10 cm, when scanning
is carried out from a flying altitude of 400 m (Baltsavias, 2000). The accuracy decreases when
the flying altitude increases (Baltsavias, 2000). Recently some manufacturers have advertised
laser implementations with a positional accuracy of 42 cm when scanning is carried out from
2000 m. The accuracy is moderate compared to traditional aerial photogrammetry. However,
the power of laser scanning is not in single accurate measurements, like in photogrammetry,
but in multiple measurements, which cover the area and can, for example, be averaged to
represent surfaces. Point density is maybe the most important laser scanning data feature and
discussed in more detail in the next section.
2.7 Point density in laser scanner data
In different airborne pulse laser systems, measurement frequency varies from a few thousand
Hz up to 150 kHz. To avoid confusion, the previous transmitted pulse has to return before the
next one is emitted, and therefore measurement frequency is limited when high flying
altitudes are applied (Batsavias, 2000). Previously flying altitudes were restricted because of
the low power of returning signals rather than confusion in pulse reception. However, some
systems today are operational from up to 6 km (www.leica-geosystems.com), and in these
measurements the frequency is limited by the time of flight of pulses. The number of
transmitted pulses does not directly correspond to the number of measured points, since
several echoes are registered from one pulse. Objects and the vertical distribution of the cover
area affect the point density.
21
Pulse distributionfunction shape Pulse distribution
function values
Shadow area
Object areaPulse amount
Scanner
Figure 2.7. Laser scanning parameters and number of pulses
The connection between laser scanning parameters and point density is characterized in
Figure 2.7. The scanning frequency and scanning mechanism determine the shape of the pulse
distribution function. In conical scanning, the amount of returns increases towards swath
edges in a non-linear way. In parallel scanning, the number of pulses decreases in a linear way
towards the swath edges. The flying altitude is a scale factor; the higher the platform flies,
the wider the swath becomes and also the distance between points increases. In the flying
direction, pulse density is affected by the velocity. In order to attain a high pulse density, the
platform should fly slowly at a low altitude.
Scanning angle
Figure 2.8. Shadow area and scanning angle
Also the scanning angle has an impact on pulse distribution. Vertical objects, which beams
are not able to penetrate, cause shadow areas behind them according to the scanning angle
(Figure 2.8). There is no shadow area immediately beneath the aircraft (scanning angle 0°).
However, near the swath edges, where pulses are transmitted in an oblique direction, shadow
areas may become large. To calculate the number of pulses reflected from a certain object, the
pulse distribution function should be integrated over the object surface projection and the
shadow area, also taking into account whether the object itself is in the shadow area of some
other object.
22
3 MATERIALS
3.1 The Otaniemi test site
In the study, five different laser scanner data sets from the Otaniemi test site were applied. This
suburban 16 hectares test site (60° 11.147' lat, 24° 49.771' lon) contains a university campus,
buildings, trees, roads and a sports field and is located in the town of Espoo, southern Finland.
The active research area was rather limited, (260 m long and 265 m wide) because it had to be
covered by all laser data sets. Elevation differences in the area are small; the elevation ranges
between 0 – 12 m. The main tree species were spruce (Picea abies), pine (Pinus sylvestris) and
birch (Betula pendula).
Figure 3.1. The Otaniemi test area; ortho photo on the left, digital crown height model in the middle, and digital
terrain l model on the right.
3.2 Laser scanner implementations
The laser measurement campaigns were carried out between 2000 and 2005 with three different
laser systems, Toposys –1, TopEye MK I and TopEye MK II.
Toposys–1 was a German-manufactured laser scanner system operational from both airplane and
helicopter. The system is no longer marketed by the company, but has been replaced with second-
generation system. A scanning mechanism was carried out through 128 light fibres in an array,
resulting in a one-directional oblique line pattern (Figure 3.2). The measurement frequency was
high, 83,000 Hz. However, point spacing in the across-track direction was significantly lower
23
than in the along-track direction. When scanning was applied, the beam ellipses overlapped in the
flying direction, whereas in the same time there was still space between adjacent scan points. The
point distance was not constant during one scan, but increased towards the strip edges
(distribution function in Figure 3.3). In Toposys-1, the scanning angle varied from +7° to -7° and
either the first or last echo was registered from one flight line.
The Swedish company TopEye has two laser scanner systems, MK I and MK II. The older
model, MK I, is operated from a helicopter and pulse distribution is carried out using a rotating
mirror. The technique produces an oblique line pattern (Figure 3.2) where the scanning angle
varies from +20° to -20°. Compared to the Toposys system, instrument measurement frequency is
low, 8000 Hz, but the advantage is a more uniform sampling pattern, in which the distance
between points in the along-track direction is close to the distance between points in the across-
track direction. In this system, the point distance along the scan increases towards the strip edges
(Figure 3.3). The first, last and even five echoes (coordinates and intensity) in the middle (if they
exist) are registered by the instrument. An additional component is a Hasselblad digital camera,
which is mounted to the system and captures images at intervals of 2.5 seconds.
The MK II system is the most recent laser scanner applied in the study. The scanning pattern is
an ellipse that rotates in the flying direction (Figure 3.2). Due to the scanning pattern, every part
in the cover area is scanned twice, first when the front edge of the ellipse crosses the object and a
second time when the back edge follows. The scanning angle varies between 14° and 20°. The
difference compared to the other systems is the shape of the point distribution function (Figure
3.3). The number of pulses is lowest just below the flight line and increases towards the edges of
the strip. In MK II, the measurement frequency is 50,000 Hz and the measurement mode is the
same as in MK I: the first, last and even five echoes and intensities are registered by the
instrument. The system can also be used in the full waveform measurement mode.
24
Figure 3.2. Toposys –1 data on the left, Topeye MK II data in the middle and Topeye MK I data on the right. An
arrow expresses the flying direction.
Poin
t den
sity
PARALLEL LINEPATTERN
Strip width
Poin
t den
sity
ELLIPTICAL PATTERN
Strip width
Figure 3.3. Point distribution function for parallel line and elliptical patterns.
3.3 Laser scanner data sets
The Otaniemi test area was scanned from altitudes 200 m and 550 m using TopEye MK I
implementation, from 400 m and 800 m altitudes using Toposys implementation and from 300 m
altitude using TopEye MK II implementation. Due to the differences in flying altitude and
systems used, the beam sizes, strip widths and average point densities varied in data sets, and
they are listed in table 3.1.
Table 3.1. Laser scanner data sets and their parameters in Otaniemi test site.
Scanner Year Flying Height Beam Size Scanning Angle Average
Point Density /m2 Strip Width
TopEye MKI 2002 200 m ~ 20 cm +-20° 2.3 110 m
TopEye MKI 2002 550 m ~ 55 cm +-20° 1 360 m
TopoSys 2000 400 m ~ 20 cm +-7° 9.9 100 m
TopoSys 2000 800 m ~ 40 cm +-7° 1.8 180 m
TopEye MKII 2004 300 m ~ 30 cm +14-20° 19.6 240 m
25
3.4 Acquisition of reference material
Tacheometer measurements were carried out in summer 2002 for two purposes: to collect ground
control points for digital terrain model evaluation and to locate tree trunks. The tacheometer was
registered to a local co-ordinate system, and afterwards the points were transformed to the
Finnish National Coordinate System (KKJ 2). The 250 ground control points were measured
along four cross sections. The points were classified into three classes according to the spatial
openness of the area, meaning access from the air, and ground cover. Classes were tree-shadowed
grass area, asphalt and grass.
The trunk location measurements were carried out as illustrated in Figure 3.4. In order to derive
the trunk centre, the direction was captured in the middle of the tree trunk, but distance
measurement was carried out to prism beside the trunk. Also these points were transformed to the
KKJ 2 system.
Direction measurement
Distancemeasurement
TRUNK
PRISM
Hypsometer
Tree height
Crown height
Figure 3.4. Tacheometer trunk location measurement. Figure 3.5. Hypsometer measurement.
In addition tree heights were recorded. Fifty trees were chosen randomly from area as test trees
including spruces, pines and birches. However, requirements of the measurement were taken into
account. The measurements were performed with a hypsometer from the distances of 20 m or 15
m. The distribution of tree heights is illustrated in Figure 3.6, and information on tree species is
given in table 3.2. Also crown depth was determined as follows: the height of the branch-free
area was measured and reduced from the tree height. However, in many cases the distinction
between branch and non-branch parts was difficult, even impossible, as the trees were not
symmetrical, and therefore the accuracy of this method was low. According to the official
26
definition of lower boundary of crown, single braches, which are below two dead year branches
are not included to the crown.
In addition to tacheometer and hypsometer measurements, the side view imaging of trees was
carried out during the field measurements. The motivation for this new technique and its method
and workflow are described in more detail in the next chapter.
0
5
10
15
20
25
30
35
12 14 16 18 20 22 24 26 28 30 32
Height (m)
fi/n,
%
Figure 3.6. Distribution of tree heights; fi is frequency
in class and n total number of samples.
Table 3.2. Tree species.
Tree species Number of
samples
Birch (Betula pendula) 24
Spruce (Picea abies) 10
Pine (Pinus sylvestris) 15
27
4 SIDE VIEW IMAGES
4.1 Motivation
A traditional forest inventory applies field instruments such as hypsometers, relascope,
bussole, measurement scissors, tacheometer, and laser pointer in order to acquire information
about tree geometry. Inventory is typically carried out in systematically located circular plots
where measurement takes place in the plot center. Single tree trunk locations are determined
with respect to the plot center, whose location is measured for example with a GPS or
tacheometer, or if the inventory has taken place previously, derived from a geographic
information system (GIS) database or a tree location map. Tree heights are acquired with
hypsometer and trunk diameters at breast height are obtained by measurement scissors.
However, acquiring the crown shape by traditional methods is problematic: if tacheometer
measurements are applied, the prism should be moved around the crown sides, not just near
the lowest branches but also to the tips of the largest branches. Another approach is to apply
measurement with trigonometric methods; the tips of the branches are projected visually to
the ground level and their locations are determined. The elevations of the branch tips are
measured separately with vertical angles. Unfortunately both of these approaches are very
laborious and time-consuming. Therefore a method of side view imaging of trees was created
to collect reference material from crown shapes.
In this method, images of trees are captured with digital camera from a terrestrial platform, a
tripod, using determined imaging geometry. Afterwards images are rectified to a plane
parallel to the tree trunk. Crown profiles are acquired in the direction of the image plane. The
idea behind this approach is the assumption that the tree trunk is vertical and branches open
out from the trunk in all directions. It is also expected that the length of the branches in
different height layers does not vary and therefore branches growing towards the camera are
not shadowing the side profile captured by the central projection of the image. Tree heights,
which are traditionally measured by hypsometer, can also be extracted from the images.
28
4.2 Capturing side view images at the Otaniemi test site
The terrestrial images were captured from the Otaniemi test area trees in summer 2003.
Imaging was done with a Nikon E-10 digital camera mounted to a camera platform with
flexible elevation and rotation possibilities. The maximum resolution, 3008×2000 pixels in
image frame, was utilized and the size of the camera aperture was fixed. During the image
capture the camera was manually focused to eternity, like in camera calibration. The same
imaging geometry was applied for all camera poses for the reason that image rectification
parameters would be solved only once and applied to all images as an automatic batch
process.
Before the actual tree imaging took place, several tests were carried out in the field to find the
optimal image capture workflow. For the measurements of crown dimensions both the tree
base and the tree crown top should be visible in the same image. However, images had to be
captured as close as possible in order to do detailed measurement. During the tests it was
observed that even in sparse forests, other trees or other objects typically shadowed the target
tree if the image shooting distance was too long. The camera stand was approximately 1.5
meters high and imaged trees were between 17-27 meters in height. In the ground co-ordinate
system -axis was towards magnetic north, -axis was in gravity direction and -axis
was horizontal and perpendicular towards - and -axes. Respectively, the parameters κ, φ
and ω are rotations between the ground co-ordinate system and the camera co-ordinate
system. The applied tests and experiences resulted in the following image capture workflow:
gz gy gx
gz gy
The image capturing location was chosen at a range of 20-30 meters from the tree.
The camera stand was adjusted according to a bubble level.
The angle κ was set to 90° using a tube level (portrait image).
A reference scale bar was attached to the tree trunk.
The camera was rotated to a position in which trunk was in the middle of the image.
ω was set to 20°, angle was measured using hypsometer.
After capturing the image, the angle φ, which is the azimuth, was measured with a
compass.
The distance to the tree was acquired with a measurement tape.
29
Figure 4.1. Image capture. From left to right: image was taken in a way that the trunk was in the middle of the
image, a reference scale bar was attached to tree trunk, the camera was tilted 20°, the distance to the tree was
measured.
In order to determine crown profiles from more than one direction, an attempt was made to
image the same trees from another direction perpendicular to the first one. Unfortunately, as a
consequence of limited free viewing directions, the image optical axes were seldom at an
angle of 90°. Eventually 15 trees were imaged from one direction and 15 others from two
locations.
4.3 Image preprocessing
In the image preprocessing, lens and sensor distortion errors were removed by applying the
camera calibration. The correction caused the image area to curve towards the image center
from the edges (Figure 4.3, middle). The calibration was carried out for the shortest focal
length f utilized in image capturing. Next, images where rectified to a plane parallel to the
trunk (Figure 4.2).
ω=
Optical axisTrunk
Rectification plane
φ
Figure 4.2. Imaging geometry and rectification plane. The side view (left) and top view (right).
30
Locations of the image corner co-ordinates on the rectification plane were calculated based on
the camera opening angles and image capture geometry. These co-ordinates and traditional
least square adjustment were applied in order to determine eight parameters of projective
transformation from image plane to rectification plane, which was carried out to all side view
images (Mikhail et al, 2001) (Formulas 4.1 and 4.2).
100
111
++++
=yfxe
gyfxeX and
100
222
++++
=yfxe
gyfxeY (4.1)
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡=
⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥⎥
⎦
⎤
⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢⎢
⎣
⎡
⎥⎥⎥
⎦
⎤
⎢⎢⎢
⎣
⎡−−−
−−
......1000
0100
2
1
2
1
0
2
1
0
YX
ggfffeee
xYyxYxxXyxXx
(4.2)
where x, y image coordinates,
X, Y co-ordinates on the rectification plane and
unknown parameters 21210210 ,,,,,,, ggfffeee
Figure 4.3. Original image (left), distortion free image (middle) and rectified image (right).
As a result of the rectification process, objects were geometrically correctly projected on the
rectification plane. The plane was vertical and in the same distance as the tree from the image
capturing position (Figures 4.2 and 4.3, right).
4.4 Image measurement
The image measurements took place in a frame where the origin was located at the tree base,
the tree height determined the maximum vertical dimension, and the maximum crown width
31
determined the horizontal dimensions (Figure 4.4). A scale parameter to transform pixel units
to metric units was obtained from a reference scale bar attached to the tree trunk. Next, tree
crown shape was determined by measuring crown width in horizontal direction. The
measurements were accomplished with 1 m intervals. In addition to the crown profile,
maximum crown width, i.e., the width of the crown projected to the ground, and tree height
were determined from the images.
Tree frame
Horizontal parallel lines
Tree height
Scale barOrigo
Maximum crown width
Figure 4.4. The tree height and crown widths were measured from two images having different viewing
directions.
4.5 Determining the trunk tilt
In forest inventory aided by remote sensing methods, tree trunks are typically assumed to be
straight. Also with instruments used in field measurement, such as hypsometers, the
assumption is the same. The method based on side view images, however, enables the
possibility of determining trunk tilt in cases where at least two tree images are captures from
different directions and the relative orientation of the images is known. Trunk tilt
determination was tested in this study.
The tree top shift sX, i.e., the horizontal deviation from the vertical line initiated from the tree
base, was determined in both images. The trunk was represented as a vector (Figure 4.5,
Formula 4.3)
kcjbiav ++= (4.3)
kcvz = (4.4)
32
where i , j and k are unit vectors in the directions of co-ordinate system axes, a and b are
unknown coefficients and c is the vertical component of the trunk vector. In the presentation
of the vertical part of the trunk vector kcvz = (Formula 4.4), c is the height of the tree. The
unknown parameters a and b are solved from an equation pair (Formula 4.5) where a and bb
are projections of shifts sX1 and sX2 to the co-ordinate system axis (Formula 4.6):
babbbabaaa
−=+= (4.5)
)2/cos(2
2 πφ −=
sXaa , )cos(
1
1φsXbb = , (4.6)
The parameters ab and ba are calculated from angles between image planes and parameters a
and b (Figure 4.5) )2/tan( 2 πφ −⋅= bab and )tan( 1φ⋅= aba . Meanwhile the trunk tilt angles
towards the image planes were calculated from vector components a, b, and c and angles 1φ
and 2φ .
φ1
φ2
aa ab
a
b
bb
ba
IMAGEPLANE 1
IMAGEPLANE 2
sX1
sX2
v
IMAGEPLANE
cv
Figure 4.5. Top view, vector projected to the ground plane (left). Side view, vector projected to the vertical plane
(right).
As expressed in Formula 4.6, determining trunk tilt requires information about the tree top
shift from both images. However, tree crowns may be wide at the top, particularly in
deciduous trees, which makes it difficult to locate tops and decreases the accuracy of
determination of trunk tilt. In addition, it is assumed in the calculation that even if the tree
trunk is in an oblique position, it is still straight. In nature this is generally not true.
Considering these limitations, trunk tilt was not determined for all trees studies in the thesis,
but applied as additional information when the results were analyzed.
33
4.6 Error analysis
In the study, the side view image measurements were compared to laser-derived vector model
features. The assumption was that image measurements were more accurate and they would
be applied as reference material. Therefore, image measurement accuracy was considered
with respect to two major error sources, scale and rotation determination errors. A simulation
was carried out to visualise the error.
During the field measurements rotations ω, φ and κ were determined with simple instruments
that have a moderate level of accuracy, namely a hypsometer, a compass, and a tube level.
The rotation ω, which is around -axis, was measured with a hypsometer. The accuracy of
hypsometer rotation measurement was estimated to be
gx
+ 1° based on the instrument ruler
index and field measurement experience. Since the image plane was rotated in preprocessing
by only 20°, the measurement plane would be at an angle of + 1° instead of in a vertical
position during the determination of crown shape.
The influence of error was simulated by rotating the image plane 1° and –1 ° around the
bottom edge of the image and then calculating the shift of each pixel compared to the correct
image. The tendency towards error behaviour is expressed in Figure 4.6. In both rotations the
effect is largest at the image edges and corners and smallest near the center of the image,
where trees were located in the capture process.
Figure 4.6. Error surfaces. Pixel dx shift for +1° ω error (up left). Pixel dy shift for +1° ω error (up right). Pixel
dx shift for -1° ω error (bottom left). Pixel dy shift for -1° ω error (bottom right).
34
The maximum errors were calculated from surface values according to Formula 4.7:
)max(2 dxDx ⋅= )max()min( dydyDy += (4.7)
)( dcfDs+
= (4.8)
D distance between projection centre and target
(f+dc) distance between projection centre and rectification plane
The maximum errors are illustrated in Table 4.1. In order to compare values with other data
sources, errors were transformed from pixels to metric values. A scale s was expressed as the
ratio of the distance between projection centre and rectification plane and the distance
between projection centre and target (Formula 4.8). During the field measurement, the image
capture distance was typically 25 meters, but all images were taken from a distance of less
than 30 meters. Maximum errors in meters for a distance of 30 m are calculated in Table 4.1.
The rotation κ (around -axis) was adjusted to 90° (portrait image) with a tube level. The
maximum error in rotation measurement was estimated to be
gz
+ 2°. Also in this simulation, the
error was considered for the worst case, which takes place when dimension measurement is
carried out from image edge to edge. For an image with a size of 3945×2928 pixels, the errors
are:
8,1)2cos(/292829284,2)2cos(/39453945
−≈−=−≈−=
dydx
The maximum errors in metric values for a distance of 30 m are calculated in Table 4.1. In
practice, trees typically filled only half or one third of the image frame and therefore the
errors were respectively smaller.
35
Table 4.1. Maximum errors.
Rotation angle Error type Error in pixels Error for distance 30 m
ω=1° dx 62 0.48 m
ω=1° dy 48 0.38 m
ω=-1° dx -62 -0.50 m
ω=-1° dy 46 -0.37 m
κ=2° dx -2.4 -0.002 m
κ=2° dy -1.8 0.014 m
κ=-2° dx -2.4 -0.002 m
κ=-2° dy -1.8 0.014 m
The optical axis of camera was adjusted in the capture process straight towards the tree trunk.
The rotation φ, which is around -axis, was measured with a compass. This method was
perhaps the most inaccurate of all methods of rotation determination, since the compass is
affected by, for example, the presence of power lines or magnetic targets in the area. The
declination between the compass north and the Finnish local co-ordinate system vertical axis
is 6° in southern Finland. However, the error in φ measurement does not affect the precision
of image measurements. The projection of objects on the rectification plane, perpendicular to
the imaging direction, are geometrically correct. The effect of error can only be noticed when
measurements are compared to other data sources. Trees with a symmetric trunk would look
the same from all directions and the error would be minimal. Unfortunately tree shape is
typically complex. The final effect of the error in φ is therefore dependent on the measured
object itself, which is beyond the scope of simulations.
gy
Finally, the error in scale measurement was considered. Images are result of perspective
projection, and therefore objects at different depths have different scales. The determined
scale s is correct only for objects located on rectification plane like the scale bar. It is assumed
that the tree crown is widest in the rectification plane. However, tree crowns are irregular in
shape. Typically the image capture distance was 25 m and the crown width was 5 m.
Therefore the distance from the crown to the camera varies from 22.5 to 27.5 m. Within this
interval the change of the scale is 0.006 to 0.0073. For example if crown width measured
from the image is 1200 pixels and its real distance to the camera is 24 m instead of 25 m, the
error in dimension is 36 cm.
36
When all error sources are taken into account, the general accuracy of this method was found
to be better than the accuracy of laser-derived vector models. The calculations in this chapter
were carried out for the worst scenario, which is not the typical situation. Therefore this
method was found to be suitable for reference material. The accuracy of this method could be
improved with more accurate rotation measurement.
4.7 Method analysis
In this chapter, a method based on side view images is suggested for measuring tree crown
dimensions. The advantage of the method is that it does not require expensive instruments,
such as terrestrial laser scanner or real-time-kinematic GPS, but relatively cheap instruments
like digital camera, camera platform, hypsometer, compass, and reference scale bar can be
applied. Collecting reference data with this method is also faster than with the traditional
method. The automation level of image preprocessing is high, because the relative orientation
between the image plane and the rectification plane is constant. Image rectification is
therefore a batch process. Image rectification does not require known points in the image area
or targeting.
However, the method based on side view images also has several limitations. The basic
problem is whether imaging from two directions is enough for crown modelling. Even if the
images are captured from two directions, the measurements are only two-dimensional. The
reason for not applying the traditional 3D stereo imaging procedure is the purpose to keep
imaging simple and efficient. Stereo measurements would require measuring corresponding
points for determining relative orientation and control points for exterior orientation. With an
irregular tree crown, these would be difficult to find. The level of tree crown shape
reconstruction may be increased if images are captured from more than two directions.
37
5 THE VECTOR MODEL PRODUCTION METHOD
This chapter introduces a method for producing vector models from original laser points and
presents features extracted from these models. The workflow steps are (1) classification of
laser points, (2) extraction of laser points from single trees, (3) vector model creation and (4)
feature calculation process. In order to acquire information about tree height, a digital
elevation model of the forest floor is required. The DTM of the cover area may already be
available from another source, such as a national elevation model or preferably derived from
laser scanning data, but since errors related to the DTM have an effect on values of the
extracted tree heights, DTM production is described as one workflow step in the thesis. For
the purpose of automating the whole process, also the issue of segmentation is considered.
The workflow is as described in Figure 5.1.
5.1 Preprocessing of airborne laser data
The preprocessing of airborne laser data is typically carried out by the company in charge of
the laser measurement campaign. Laser ranges, DGPS and IMU measurements are combined
based on a time stamp attached to each data source, and co-ordinates are calculated for laser
pulse reflection points and digital camera image capture locations. However, this so-called
direct orientation approach is affected by many errors, such as improper satellite geometry
during the laser scanning, calibration errors, and misalignment of the laser scanner and digital
camera. A strip adjustment is carried out in order to match data from adjacent laser strips and
to reduce the impact of these errors. The matching process typically applies data from
calibration areas, which are flown repeatedly during the scanning. A detailed description of
matching is found in Burman (2000). Since measurements are in the global co-ordinate
system WGS84, they are transformed to the local co-ordinate system afterwards.
In this thesis, the local co-ordinate system was the Finnish KKJ 2, and the laser data utilized
in the study was transformed to the system by applying one of the two different WGS84 –
KKJ 2 transformations. The Finnish Geodetic Institute has published one transformation
program based on equations from Ollikainen (Ollikainen, 1993). Also the National Land
Survey of Finland runs a program to transform co-ordinates from WGS 84 to KKJ either in
northern or southern Finland based on equations published in JHS 154 (JHS 154). The
transformation based on Ollikainen’s model is considered to be more accurate. Therefore
38
results from the National Land Survey transformation were adjusted to match that
transformation model. In the Otaniemi area, the difference is a simple constant shift in the x-
and y- directions; dx=0.641 and dy=0.268. The geoid fluctuates in the Otaniemi area between
–18.062 and 18.054 m (dz=0.8 cm). The variation was considered to be minor and therefore
constant shift was applied to z- co-ordinates.
Preprocessing
Classification
Digital terrain model
Digital environmental model
Digital tree height model
Segmentation surface
Tree point extraction
Vector model creation
Original points
Figure 5.1. Vector model workflow.
5.2 Point classification and DTM
Several returns are typically obtained from one emitted pulse in forested areas, since part of
the beam is reflected from the tree canopy and other parts from the forest floor. Typically the
first returns are reflections from objects close to the platform, i.e., the forest canopy, and
correspondingly the last returns are reflections from the forest floor, in case the beam has
been able to penetrate through the canopy. However, first and last echoes occasionally
originate from both ground and vegetation, which is why a classification algorithm is applied
to separate these. The common approach in classification is to determine ground points first,
since the ground is continuous surface with a smoothly varying altitude. Afterwards,
vegetation points are classified based on the distance from the ground.
Several algorithms have been developed to obtain a DTM from laser scanning point clouds.
Kraus and Pfeifer (1998) developed a DTM algorithm based on distinguishing laser points
into terrain points and non-terrain points using an iterative prediction of the DTM and weights
attached to each laser point depending on the vertical distance between the expected DTM
39
level and the corresponding laser point. Pyysalo (2000) developed a modified recursive
classification method for DTM extraction, where all points within a vertical distance of 60 cm
from the lowest expected ground level were included equally in the next DTM model
calculation. Elmqvist (2001) estimated the ground surface by employing active shape models
by means of energy minimization. The active shape model behaves like a membrane floating
up from underneath the data points. The energy function is a weighted combination of internal
and external forces. The start state is a plane below the lowest point in the data set. Sithole
(2001) and Vosselman and Maas (2001) developed a slope-based filtering technique, which
works by pushing a structuring element up vertically. In the method used by Wack (2002),
non-terrain raster elements are detected in a hierarchical approach that is loosely based on a
block-minimum algorithm.
The method applied in this thesis comes from Axelsson and has been implemented in
Terrascan software. Axelsson (1999, 2000, 2001) developed a progressive TIN densification
method where the surface was allowed to fluctuate within certain values controlled by
minimum description length, constrained spline functions, and active contour models for
elevation differences. Ground points were connected in a TIN. A sparse TIN was derived
from neighbourhood minima, and then progressively densified to the laser point cloud. During
every iteration round, points are added to the TIN, if they fall within the defined thresholds.
The algorithm was applied to create five different DTM, one from each data set. However,
before ground classification took place, low points, which are multi-path reflections, were
separated to avoid their confusing effect on algorithm behaviour. The classification was
carried out by applying parameters illustrated in table 5.1. The result was that approximately
30 % of points were defined as ground. The DTM was exported as a grid model with a pixel
size of 1 m (Figure 5.3).
Table 5.1 Ground classification parameters.
Maximum building size 60 m
Maximum terrain angle 80°
Maximum iteration angle to plane 6°
Maximum iteration distance to plane 1.4 m
Iteration angle was reduced when edge length was less than 5 m
40
Figure 5.2. Original points in TopEye MK I data. Figure 5.3. Exported DTM.
5.3 Tree crown delineation and point extraction
In the next workflow step, tree point extraction, the reflections from each tree are
geometrically separated from surrounding points. Two approaches were applied in this thesis,
namely the manual and segmentation-based approach. In both approaches, additional surfaces
were required, a digital surface model (DSM) and a canopy height model (CHM). In the DSM
each surface cell obtained its value from the highest measurement within pixel (Figure 5.4).
Altitudes were relative to the sea level. The CHM was obtained by reducing the DTM from
the DSM. In the resulting CHM, elevations were from the ground surface (Figure 5.5).
Figure 5.4. Digital surface model (DSM). Figure 5.5. Canopy height model (CHM).
In the manual extraction approach, vector polygons were drawn by hand around each tree
(Figure 5.6). The CHM was applied below the vector platform (MicroStation) in order to
visually guide delineation. After this, laser points within each vector element were exported to
separate files (TerraScan) from all different laser data sets.
41
Figure 5.6. Manual tree crown delineation
The segmentation approach performed automatic delineation of tree crowns. Segmentation is
an advanced image processing method, which has been introduced in forest inventory and
applied from large-scale aerial images up to high-resolution satellite images. The methods
used in laser scanning have been applied in similar studies using aerial imagery with a very
high resolution, with the distinction that the image is replaced by the crown DSM or the
CHM. For operational forest inventory automatic segmentation is a minimum requirement,
since the manual approach is too time-consuming.
Several algorithms have been developed for automatic tree crown segmentation. According to
Gougeon and Leckie 2003, two parts of the process are tree detection and full crown
delineation. Tree locations can be found, for example, by detecting image local maxima (e.g.
Geogeon and Moore 1989). The method provides that the filter size and image smoothing
parameters are appropriate for the tree size and image resolution. With that assumption the
approach works relatively well with coniferous trees (see also Gougeon and Leckie, 2003).
After finding the local maxima, the edge of the crown can be found using the processed
canopy height model.
Hyyppä and Inkinen (1999) were the first to demonstrate laser scanner aided forest inventory
by finding maxima of the Digital Tree Height Model (DTHM) and applying segmentation for
edge detection. The method was tested together with two other segmentation algorithms in
Finnish, Austrian, and German coniferous forests, and 40 to 50 % of the trees could be
correctly segmented. Persson et al. (2002) improved the crown delineation and could link
71% of the tree heights with the reference trees. Other attempts to use DSM or CHM for
individual tree crown isolation or crown diameter estimation have been reported by e.g.
Brandtberg et al. (2003), Leckie et al. (2003), Straub (2003), and Popescu et al. (2003).
42
Andersen et al. (2002) proposed to fit ellipsoid crown models in a Bayesian framework to the
point cloud. Morsdorf et al. (2003) presented a practical two-stage procedure where tree
locations were defined using the DSM and local maxima and crown delineation was
performed using k-means clustering in the three-dimensional point cloud. Wack et al. (2003)
first calculated the canopy height corresponding to all laser points and used the sorted list to
define a cone of the tree top. If a point was located close to the cone, it was removed. The
process was recursive. This method allowed 93 % of planted eucalyptus trees to be correctly
delineated.
The segmentation method applied in the study was the one tested by Hyyppä and Inkinen in
1999. A detailed description is found in Hyyppä & Inkinen (1999). The crown delineation is
based on a watershed algorithm that turns the CHM upside down, picks up the local minima
(tree tops) and examines the surrounding of the minima with the help of the direction and size
of the gradient. A drop of water falling on a surface follows a descending path and eventually
reaches a minimum. The pixels in which flooding ends at the same minima are included in the
segment (Figure 5.7). Because segmentation is carried out to the surface instead of the
original 3D points, the resulting areas do not overlap. In this study, the segmentation was
carried out to all five CHMs produced from different laser materials.
Figure 5.7. Segmentation surface.
5.4 Vector model processing method
In the method, each individual tree crown is bounded by parallel horizontal polygons, and the
tree trunk is formed as a vertically oriented vector from the ground surface to the tree top. The
topology information required by the method is crown/non-crown classification and polar co-
ordinates with respect to the trunk location. In the study, applied trunk locations were
43
measured by tacheometer. Another possibility is to derive trunk locations from laser point
cloud.
A crown/non-crown classification is carried out by analysing the distribution of the z co-
ordinate histogram. The reflections from ground and undervegetation can be typically
recognized as clearly distinguished maximum in the histogram (Figure 5.8), which is followed
by a minimum. The shape of the histogram is partly due to laser scanning principle and partly
due to typical tree geometry. In airborne laser scanning, pulses approach the tree from above,
and therefore the highest laser energy is applied and received from upper canopy. On the
other hand, tree branch biomass is typically less near the ground and therefore pulses have
less volume to reflect from. The lower boundary of the crown is examined by taking
histogram minima into consideration and determining the highest altitude of these minima as
the lower boundary of the crown. Points above the lower boundary of the crown are classified
as crown points and trunk location and are calculated by using the mean values of these
points.
tx ty
Figure 5.8. Three trees and their z co-ordinate histograms; on the left a birch, in the middle a pine and on the
right a spruce.
The transformation of crown points from the x, y, z co-ordinate system to the polar co-
ordinate (α,r,h) system is carried out by applying the formulas:
44
rzzh −= , ( ) ( )22tt yyxxr −+−= ,
( )r
xx t−= arccosα
(4.1)
where h point elevation from the ground
ground altitude rz
trunk co-ordinates tt yx ,
r range between trunk and point
α azimuth
The space between the lower boundary of the crown and the tree top is split in altitude layers
of 1 m in the z-direction, and points within a particular height layer are taken into account and
organized according to the ascending angle azimuth. Each point is connected to the previous
and next point with a line according to the ascending azimuth. This procedure results in a
polygon and goes through every point, i.e., polygon node, without its side edges crossing each
other (Figure 5.9). The polygon is formed on the tree model according to the layer height. The
forming and drawing of polygons is repeated for all the crown layers (Figure 5.10). Finally, to
improve the visual outlook of model, the trunk is described in straight lines from the top of
the tree to the surface of the DTM.
The vector models of 50 trees in the Otaniemi test area, which were delineated manually,
were produced from five laser scanner data sets, resulting in 250 models. However, only one
segmentation surface, CHM from Topeye 200 data, was used in automatic delineation for all
data sets in order to avoid the impact of segmentation errors on feature extraction, as
described in the following section.
45
Figure 5.9. Polygon formed with respect of trunk location
Figure 5.10. Original points (left) and formed vector model (right).
5.5 Tree model feature extraction
The formed vector models are used to provide individual tree features, which are (1) tree
height, (2) crown height, (3) location of the trunk and (4) crown shape determined by
polygons.
The tree height is calculated as the difference between the z co-ordinate of the highest crown
point and the ground altitude at trunk location. In order to get the ground height, DTM is
applied. Crown height is obtained as the difference between tree height and the height of the
lower boundary of the canopy. In order to derive trunk location two methods are possible,
either calculate the arithmetical mean of crown points or use the plane co-ordinates of highest
measurement point, which is also used in tree height extraction.
The crown polygons are utilized to provide side profiles from a given direction. The whole
crown model is split by a vertical plane, which is perpendicular towards the given direction,
46
i.e. the image capture direction (Figure 5.12). The direction measured with a compass and
trunk location co-ordinates are used to determine the cross-section. The co-ordinates for the
two intersecting points of the plane and polygons and are calculated
based on polygon edge line equations
),( 11 cc yx ),( 22 cc yx
ee bxay += , which are derived from edge points, and
the plane equation. Finally polygon widths dw1 and dw2 are calculated as the distance
between previously determined intersection points and trunk location (Figure 5.11). An
example of profiles and original points is illustrated in Figure 5.12. The profiles are formed
from crossing points.
Cross-section
dx 2
dx 1
Image plane
Figure 5.11. A vector polygon (dotted line) and
cross-section (solid line) with respect to the trunk
location (a dot). The values dw1 and dw2 are the
widths of the crown measurements.
5.12. Example of a crown profile. The red lines are
crown widths derived from intersection of polygons
and plane equation.
In this study DTM derived from Topeye 200 data set was used in the feature extraction of all
laser-derived vector models in order to avoid the impact of DTM errors and to allow the
comparison of laser data sets. The extraction was carried out on 250 trees delineated manually
and approximate 160 trees, which were detected correctly in segmentation.
47
6 RESULTS AND THEIR ANALYSIS In order to analyse the accuracy of vector models created with the method described in
chapter 5, model features were compared with the reference materials as described here. First,
altitude values from the digital surface model were compared with the tacheometer ground
control measurements, since errors in terrain model values produce errors in tree height
estimation. Next, the extraction of pulses reflected from individual trees was considered by
analyzing tree detection in segmentation surfaces and comparing them to manually delineated
tree polygons. Then tree model heights and crown model heights were compared with the
hypsometer measurements and values extracted from side view images. The determination of
tree trunk location was considered with respect to both manual and segmentation-based tree
extraction approaches. Finally, crown profiles derived from the vector model were compared
with the side view image profiles.
0 5 10 15 20 25 30 35 40 4520
25
30PROFILES 1−4
Z (
m)
Laser DTM Tacheometer DTM
0 5 10 15 20 25 30 3520
25
30
Z (
m)
0 5 10 15 20 25 30 3515
20
25
Z (
m)
0 5 10 15 20 25 3020
22
24
Measured break point (no.)
Z (
m)
Figure 6.1. Terrain profile comparisons. The length of profiles varies between 80-122 m.
The laser scanner data sets are listed in the next tables according to the laser system
implementation, which is followed by the flying height in meters.
48
6.1 The evaluation of DTM accuracy
The DTM values were compared with the tacheometer ground control measurements. Four
profiles were measured in the Otaniemi area during summer 2003. Points were classified into
three classes and outliers according to the open access from the air and ground cover type.
These classes were asphalt area, grass area, and tree-covered grass area. The classification
was carried by means of exploiting orthoimages. All five DTMs were applied. DTM profiles
from TopEye 200 data and tacheometer measured profiles are in Figure 6.1, standard
deviations and bias values in Tables 6.1 and 6.2.
Table 6.1. Standard error of differences of digital terrain model comparison.
Std (m) Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Tree-Covered-Area 0.08 0.083 0.14 0.19 0.11 Asphalt 0.036 0.026 0.07 0.05 0.035 Grass 0.046 0.047 0.09 0.06 0.07
All except outliers 0.07 0.07 0.15 0.13 0.08
Table 6.2. Bias error of differences of digital terrain model comparison.
Bias (m) Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Tree-Covered-Area 0.15 0.047 0.18 0.19 -0.015 Asphalt 0.05 -0.007 0.06 0.08 -0.05 Grass 0.07 -0.01 0.03 0.13 -0.02
All except outliers 0.08 0.0009 0.05 0.13 -0.013
The standard deviation of elevation comparison differences was 7 cm, when the DTM from
TopEye 200 data was applied. According to the laser system manufacturer, the accuracy of
distance measurement is as high as 6 cm. This results from the technical restriction of laser
pulse transmission. However, an accuracy of 6 cm may be achieved only if the whole pulse is
reflected from a plain surface perpendicular to the laser system. This is not the case on forest
floors, where pulses are likely to reflect from several undervegetation surfaces. The laser
measurement campaigns utilized in this study were carried out during the summer, when
49
forest ground was typically covered by grass and hay. Therefore, closer attention was paid to
ground cover in tacheometer measurement locations.
The control points on the asphalt-covered road and parking area had the smallest biases
(-0.007 - 0.08 m) and standard deviations (0.035-0.07 m) (Table 6.1). The accuracy decreased
in grass areas, even if surfaces were plain. The intensity of grass-reflected pulses was
approximately 20 % less than from asphalt areas, and it was visually observed that the
deviation of point altitude was higher in grass areas. Differences between ground control
measurements and the digital terrain model were highest in areas shadowed by trees, from
-0.015 m up to 0.19 m. These results suggests that in forested areas the ground surface is
generally higher in the laser-derived model than the real ground truth.
Results obtained in the study were similar to those suggested by previous studies. Kraus and
Pfeifer (1998) gained an RMSE of 57 cm using ALTM 1020 and average point spacing of 3.1
m in wooded areas. Hyyppä et al. (2000) reported a random error of 22 cm for modulating
forest terrain using Toposys-1 and nominal pulse density of 10 pulses per m2. Three different
DTM algorithms were compared within the EC-funded HIGH-SCAN project (1998-2001) in
Finnish (test site Kalkkinen), Austrian (Hohentauern) and Swiss (Zumikon) forests. Obtained
random errors varied between 22 and 40 cm (Hyyppä et al., 2001) using Toposys-1 and pulse
densities between 4 to 10 pulses per m2. Ahokas et al. (2002) compared three algorithms on a
forested hill in Finland and found random errors between 13 and 41 cm using Toposys-1.
Reutebuch et al. (2003) reported random errors of 14 cm for clear-cut forest, 14 cm for
heavily thinned forest, 18 cm for lightly thinned forest and 29 cm for uncut forest using
TopEye data with 4 pulses per m2. However, in dense forests, errors up to 10 to 20 m can
occur in the DTM estimation (Takeda, 2004).
6.2 Single tree point extraction by tree delineation
A tree crown delineation was performed in the study by using two approaches, manual and
segmentation-based, which were carried out to five CHMs obtained from laser data sets.
Results from manual delineation and automatic segmentation are given in Figures 6.2-6.7.
The manually drawn tree polygons, which were revised during the field inventory, were
considered as correct. At this stage, the location of tree crown borders were not considered,
only the detection of trees.
50
Figure 6.2. Manual segmentation. Figure 6.3. Segmentation, TopEye 200 derived CHM.
Figure 6.4. Segmentation, TopEye 300 CHM. Figure 6.5. Segmentation, TopEye 200 CHM.
Figure 6.6. Segmentation, Toposys 400 CHM. Figure 6.7. Segmentation, Toposys 800 CHM.
Figure 6.8. Manual (green) and segmented TopEye 200 CHM polygons (blue).
51
Firstly, the number of correctly detected trees was considered with relation to all trees (Table
6.3, row 1). Correct detection refers here to a segment that includes only one single tree, other
solutions were considered as errors (Table 6.3, row 2). Secondly, erroneous trees were taken
under closer study. Two types of errors took place. The first error consisted of merging more
than one tree crown in one segment. The other error, splitting, occurred when area containing
a single tree was divided into several segments. The proportion of these with respect to all
erroneous trees is presented in Table 6.3, rows 4 and 5.
Table 6.3. Tree detection errors applying automatic segmentation.
% Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Correct 65.91 65.63 63.64 34.09 36.36
Error 34.09 34.38 36.36 65.91 63.64
Merged 34.09 34.38 29.55 38.64 27.27
Splitted 0.00 0.00 6.82 27.27 36.36
The segmentation results indicated what previous studies have shown, namely that the
accuracy of automatic segmentation is moderate. The proportion of correctly detected trees
varied from 34 to 66 %, and the best results were gained with CHM derived from TopEye
data flown from an altitude of 200 m. The proportion of merged tree crowns was
approximately the same in all CHM segmentations and errors were typically related to the
same trees. However, the proportion of divided tree crowns increased with rising flying
altitude and decreasing point density. All TopEye-derived CHMs were more accurately
segmented than surfaces produced from Toposys data. The reason for this was suggested to be
the scanning pattern of TopEye implementation. Regardless of the extremely high scanning
frequency of Toposys, the point distance along the scan direction is eight times larger than the
across-scan point density. Therefore, some pixels in the CHM may result from only one laser
hit that has been reflected from 1 square meter of irregularly shaped tree crown. The system
scanning angle in Toposys is also steep, which results in laser beams reflecting from crown
tops instead of crown sides. It was speculated that crown sides were moderately presented in
the CHM.
52
In both manual and segmentation-based delineation approaches, the fundamental problem was
that tree crowns were not assumed or allowed to overlap. Short trees growing under higher
trees and trees with jointed branches were therefore automatically erroneously segmented
(Figures 6.9 and 6.10). The reason for this is that the segmentation was applied to a surface
where each pixel has only one value, and the original measurement point cloud was ignored.
Correspondingly, in manual delineation, the polygon edges split the point cloud in a vertical
direction and two points with same x and y co-ordinates may not end up in two different
groups. Therefore it was concluded that delineation accuracy originates not just from the
algorithm method but also from the forest structure itself. The space between trees that
separates them from surrounding neighbours improves detection and delineation accuracy,
whereas trees growing in tight groups are likely to be detected erroneously. Similar results
have been gained by Perrsson at al (2002).
Figure 6.9. A short tree beside a tall tree. Figure 6.10. Trees with jointed crowns.
6.3 Tree heights and crown heights
The tree height comparison was carried out first using values derived from the laser model
and hypsometer measurements and then using values derived from the laser model and side
view image measurements (Figure 6.11). The tree height was calculated as the difference
between the z co-ordinate of the highest crown point and the ground altitude at tacheometer
derived trunk location. In this study, the DTM derived from the TopEye 200 data set was
applied for all five data sets in order to avoid the effect of DTM error and to allow the
comparison of different data sets. The standard deviations and biases of residuals are
presented in Tables 6.4 and table 6.5. In the study bias values were calculated as average of
differences.
53
Hypsometer
Tree height
Crown height
Figure 6.11. Hypsometer measurements, laser model and side view image were compared.
Crown heights were obtained as the difference of tree heights and the heights of the lower
boundary of the canopy. These values were studied with respect to corresponding hypsometer
and side view image measurements (Figure 6.11). The standard deviations and biases of
differences are presented in Table 6.4 and Table 6.5.
Table 6.4. The differences of laser-derived tree heights and hypsometer measurements.
Tree height Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 1.32 1.21 1.4 1.4 1.5
Bias, m -0.44 -0.52 -0.06 -0.5 -0.005
Crown height
Std, m 4.41 5.05 4.77 5.29 3.84 Bias, m 2.68 -0.24 -1.36 -1.43 -1.92
Table 6.5. The differences of laser-derived tree heights minus and heights from side view image measurements.
Tree height Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 1.03 1.02 1.06 1.04 1.09 Bias, m -0.33 -0.06 -0.67 -0.24 -0.77 Crown height Std, m 3.76 5.38 2.98 3.20 4.64 Bias, m -0.84 -1.16 -2.48 -1.64 -3.87
In hypsometer versus laser-derived tree height comparison, the biases varied from a couple of
centimetres to half a metre (Table 6.4). As expected, all biases were negative, which indicates
that tree heights derived from laser models are smaller than those measured by hypsometer.
The reason for this assumption is that laser pulses do not necessarily reflect from the tree
54
crown top and therefore tree height is underestimated. The DTM used in the comparison was
also laser-derived, and as mentioned in section 7.1, it was typically slightly above ground
truth in a tree-shadowed area. This causes tree heights to be underestimated. The standard
deviation of residuals was 1.2 m at minimum and increased to 1.5 m as flying height
increased and point density decreased. The accuracy was slightly worse than expected.
However, the standard deviations of residuals were 30 % smaller when tree heights measured
from side view images were applied (see Table 6.5). The results were computed with all five
data sets. Field inventory experience has shown that hypsometer measurement accuracy is
only 1 m or even worse. The measurement method is disturbed by the effect of wind bending
trees and also unclear visibility of wide crown tops. Therefore, tree heights obtained from
laser measurements and side view images were potentially more accurate than hypsometer
measurements. The differences obtained from comparison of hypsometer data were also
considered tree by tree, and it was observed that some trees had large biases but small
standard deviations. This indicated that for some trees, the heights derived from different laser
data sets were similar, but differed from tree heights acquired by hypsometer. The comparison
with side view images indicated that derived tree heights are underestimated by
approximately –0.06 m - 0.77 m and with STD of 1 m. This result was consistent with
assumptions and similar studies.
A corresponding comparison was made between laser-derived crown height and hypsometer
and side view image measurements (Tables 6.4 and 6.5). In this study, crown height was
determined as the difference between the tree top elevation and the elevation of the lower
boundary of the crown, which was defined to the minimum of the z co-ordinate histogram.
The results indicated that tree crown heights were inaccurately estimated from laser-derived
tree models. High negative biases imply that laser pulses have not reflected from the lower
branches of trees, which therefore are not visible in the models. Differences also fluctuated
heavily, resulting in standard deviations of several metres. It was visually observed that tree
crowns growing in groups were particularly underestimated as laser pulses were unable to
penetrate through the canopy of neighbouring trees. However, in open areas vector models
were more complete and residuals correspondingly smaller.
55
In addition to the lack of hits on lower branches, the results were impaired by a few cases
where the algorithm failed to determine the lower boundary of the canopy. This error took
place with trees whose crown continued all the way down to the ground surface and there was
no branch-free area (see Figure 6.12). In these cases, the z co-ordinate histogram minimum
was located somewhere in the crown. With spruces lower boundary of crown was located
typically just near the top, where crown forms a peak. Modification of the algorithm could
remove this problem, but this was not done during the study, as the number of these situations
was small, only 10 out of 250. It should also be emphasized that the lower boundary of the
crown is difficult to define even with a hypsometer in field inventory and from side view
images, since in natural conditions branches do not grow symmetrically around the tree.
Figure 6.12. A spruce with a crown continuing to the ground. Side view image on the left, laser points in the
middle, and histogram of z co-ordinates on the right, where the lower boundary of the crown algorithm has
failed.
6.4 Locations of the tree trunks
The laser-derived tree trunk locations were compared to tacheometer measurements. Trunk
locations were derived by two methods; calculating the arithmetical mean of x and y co-
ordinates of crown class points and using the plane co-ordinates of highest measurement
point. Two comparisons were carried out, laser model derived values from manually extracted
trees (Tables 6.6 and 6.7) and laser model derived values from automatically segmented trees
(Tables 6.8 and 6.9) applying the CHM from TopEye 200 data. Only correctly detected trees
56
were included in the latter comparison, which is why the number of trees here was less than in
the comparison of manually delineated trunk locations.
Table 6.6. Comparison of manually delineated trees and tacheometer measurements. Number of samples in
comparison is 50 for each data set. Derivation of trunk location has been carried out calculating average of
crown hits.
distance Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 0.74 0.84 0.83 0.8 0.8
Bias, m 1.18 1.44 1.35 1.24 1.2
Table 6.7. Comparison of manually delineated trees and tacheometer measurements. Number of samples in
comparison is 50 for each data set. Derivation of trunk location has been carried out using plane co-ordinates of
the highest measurement point.
distance Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 0.79 0.91 0.83 0.99 1.43
Bias, m 1.18 1.21 1.3 1.34 1.5
Table 6.8. Comparison of segmentation-delineated trees and tacheometer measurements. Number of samples in
comparison is 32 for each data set. Derivation of trunk location has been carried out calculating average of
crown hits.
Distance Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 1.01 0.85 0.84 1.07 1.03 Bias, m 0.66 0.64 0.58 0.74 0.71
Table 6.9 Comparison of segmentation-delineated trees and tacheometer measurements. Number of samples in
comparison is 32 for each data set. Derivation of trunk location has been carried out using plane co-ordinates of
the highest measurement point.
distance Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 0.8 1 0.93 1.14 1.16 Bias, m 1.16 1.4 1.34 1.39 1.56
The results in Tables 6.6 - 6.9 indicate that trunk locations were derived approximately with
an accuracy of 2 m. In all comparisons the biases were between 0.66 – 1.56 m and standard
57
deviation of differences between 0.74 – 1.16 m. However, the average calculation method
provided locations more accurately than using the plane co-ordinates of the highest
measurement point for both manually and segmentation delineated trees.
By using the plane co-ordinates of highest point it is assumed, that highest point and trunk
location at the breast height are in vertical line, i.e. the trunk grows straight. The trunk tilt
causes the tree crown to be shifted from the root base, which is the location that tacheometer
measurement captures. Therefore trees with the tiled trunk produce errors in the comparison.
The difference in accuracy between manually or segmentation delineated trees using plane co-
ordinates of highest point was minor, since highest point was the same using both methods.
The small difference in accuracies results from the different number of trees in a sample.
The average calculation method is also affected by several factors. The method assumes that
laser pulses have been reflected from all sides of the crown, since scanning was performed
from above. In practice, the crown side towards the scanner will probably reflect more pulses
than the opposite crown side, and therefore trunk location calculated from points is shifted
towards the scanner. In addition to pulse reflection itself, two method-dependent factors
influence the accuracy. They are single tree crown delineation and point classification. An
erroneously delineated tree crown shall have an erroneously derived trunk location, if points
from neighbouring trees are included in the calculation or, in contrast, if some part of the tree
crown is ignored. In this study, two methods were applied to delineate tree crowns. As
mentioned in section 6.2, results from the automatic segmentation were moderate from the
viewpoint of tree detection. Trees that were not correctly detected in automatic segmentation
have been left out of the trunk location experiment. The correctly detected trees were,
however, located with almost the same accuracy as the manually delineated trees. The
segmentation algorithm seems to delineate the tree crown tightly around crown sides, whereas
manually delineated crowns typically included some ground area beside the tree (Figure
6.13). Even if standard deviations of segmented tree trunks were 15 % larger, the biases were
smaller.
58
Figure 6.13. Examples of tree crown delineation with manual and segmentation approaches. Red points represent
the manual approach and cyan points the segmentation approach.
The classification of points as crown points also had an impact on the trunk location
determination. Points above the lower boundary of the crown were considered as crown
points. Section 6.3. discussed the problem of detecting the lower boundary of a crown
continuing to the ground and its effect on trunk location. The erroneously detected lower
boundary resulted in crown points being ignored in the derivation of the vertical trunk
location. However, the number of trees with this problem was rather small.
6.5 Crown width analysis
The crown widths were analyzed using two methods. At first, the maximum width of crown
acquired from images were compared to maximum crown widths derived from laser models
in the image plane direction (Figure 6.14). This phase was carried out for both image capture
directions for those trees for which a second image had been acquired. Results of the
comparison of dw1 and dw2 are listed in Tables 6.10 and 6.11. Some images were left out of
the calculations because of wind bending trees at the moment of image capture or tree fusing
merging with background.
59
Maximum crown width
dw1 dw2
Figure 6.14. Maximum crown width.
Secondly, crown profiles derived from both images and laser models were overlaid and
examined visually. The crown widths were plotted as a function of layer height with respect
to the total tree height and with respect to the viewing direction. The reason for applying the
tacheometer-acquired trunk location was to avoid the contribution of the laser-derived trunk
location error. Examples of superimposing are presented in Figures 6.15, 6.16 and 6.17.
Table 6.10. The maximum width of crown projections compared to side view image measurements.
Image 1, dw 1 Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 1.57 1.51 1.44 1.55 1.38 Bias, m -0.19 0.46 -0.53 0.08 -0.52
Image 1, dw 2
Std, m 1.75 2.20 1.58 1.66 1.52 Bias, m -0.36 -0.06 -0.62 0.11 -0.73
60
Table 6.11. The maximum width of crown projections compared to side view image measurements.
Image 2, dw 1 Topeye 200 Topeye 300 Topeye 550 Toposys 400 Toposys 800
Std, m 2.20 2.48 2.27 1.99 2.00 Bias, m -0.28 -0.06 -0.51 -0.05 -0.85
Image 2, dw 2
Std, m 1.87 2.04 2.30 2.04 2.16 Bias, m -0.70 -0.61 -0.72 -0.55 -0.78
Figure 6.15. A spruce. The solid line is the image-derived crown profile and the dotted lines are crown profiles
derived from laser models. From left to right: TopEye 200, TopEye 300, TopEye 550, Toposys 400, Toposys
800.
Figure 6.16 A pine. The solid line is the image-derived crown profile and the dotted lines are crown profiles
derived from laser models. From left to right: TopEye 200, TopEye 300, TopEye 550, Toposys 400, Toposys
800.
61
Figure 6.17 A birch. The solid line is the image-derived crown profile and the dotted lines are crown profiles
derived from laser models. From left to right: TopEye 200, TopEye 300, TopEye 550, Toposys 400, Toposys
800.
As the results in Table 6.10 and 6.11 show, crown projection widths were derived from laser
models with an accuracy of approximately 2 m. The results were attained with both first and
second image capture locations. In almost all comparisons the biases were negative, which
indicates that the tree crown is underestimated in laser scanner derived models. This result
was consistent with tree height estimation, where biases were also negative, as well as
previous studies by Rönnholm in 2005 and Gaveau et al. in 2003. Also visual comparisons of
crown profiles showed the tree crown to be diminished in laser pulse derived vector polygons.
In most examples, the laser-acquired crown profile was within the side view image crown
shape.
The gained results were interpreted as indicating, that in addition of the uncertainty related to
laser pulses not hitting the top of the crown or tip of the branch, the algorithm registering laser
pulses tends to determine reflection inside the crown rather than on the surface. A vegetation-
originated pulse consists of the sum of multiple small surface reflections and the return signal
is broad and low instead of a clear strong peak. The recent development of laser scanner
implementations that register full waveforms has provided the opportunity to see the signal
shape of the vegetation surface. In the study a data set measured with full waveform mode in
test area in Remmingtorp, Sweden was examined. Example of two return signal that has taken
place on the side of a tree crown is shown in Figure 6.18. The return areas are wide and
smooth. For comparison, two reflections from a sand-covered road surface are shown in
Figure 6.19. The algorithms may possibly determine the pulse as returned in several positions
depending on the algorithm. With a quantisation interval of 14.4 cm, errors could easily
become large even with a small change in histogram shape.
62
Figure 6.18 Vegetation signal examples. Figure 6.19 Road signal examples.
A tree delineation had impact on the crown width analysis. An example of the phenomena is
presented in Figure 6.20, where crown points from neighbouring trees have been located
within the delineation polygon, and therefore the profile derived from a laser model is wider
than the image-derived profile. This example indicates that delineation affects the horizontal
accuracy of tree geometry. Another type of error took place when the wind bent the tree
during the moment of the image capture. Figure 6.21 illustrates a birch whose laser-derived
profiles had a similar shape, which, however, was different from the image-derived profile.
Figure 6.20. Erroneously delineated spruce. The
solid line is the image-derived crown profile and
the dotted line is the crown profile derived from
laser models.
Figure 6.21. Example of a tree that has been bent by
the wind during the image capture moment. The
solid line is the image-derived crown profile and
the dotted lines are crown profiles derived from
laser models.
63
7 THE IMPACT OF LASER SCANNING PARAMETERS ON THE
APPLICATION OF VECTOR MODELS
The influence of laser scanning parameters on the modelling accuracy was taken into
consideration. Laser scanning parameters are, for example, point density, pulse transmitting
angle, and size and length of the beam. In the data sets used in the study, these parameters
had several values (Table 3.1). However, the accuracy gained from tree and crown height
comparison in addition to the estimation of trunk location was approximately the same with
all five laser data sets. The impact of each laser scanning parameter was difficult to estimate
by comparing different data results, since more than one value changed between each data
pair. In order to estimate the influence of parameters on the vector models, closer attention
was paid to the data structure and data simulation was carried out to estimate the effect of the
beam size and overlapping strips.
7.1 Point density and reconstruction
The largest difference between data sets was in point density, which varied between data sets
from 1 /m2 up to 20 / m2. The point density is a result of the measurement frequency, scanning
mechanism, and flying altitude, but the amount of pulses reflected from one object is also
affected by the geometry of the measured object, scanning angle, and object penetration. Point
density is not unique, but varies according to the pulse distribution function. Examples of
number of pulses reflected from a single tree are presented in Table 7.1 and an illustration of
one pine tree is provided in Figure 7.1. The number of pulses reflected from one tree varied
dramatically depending on the data set, tree shape itself, and location in the scanning strip.
Table 7.1. Number of pulses from single example trees in different data sets.
Number of hits Birch Spruce Pine Topeye 200 540 408 201 Topeye 300 2031 715 2440 Topeye 550 169 84 95 Toposys 400 1455 688 813 Toposys 800 248 114 154
64
Figure 7.1. From left to right: Toposys 400, Toposys 800, Topeye MK I 200, Topeye MK II 300 and Topeye
550. The single pine is in the Otaniemi area.
The high pulse amounts in trees extracted from Topeye 300 data are caused by the scanning
pattern, where each object is measured twice, but also from an oblique scanning angle. Pulses
from the shadow area of an impenetrable vertical object are reflected from the object surface,
which increases the amount. The histograms of the z co-ordinate are presented in Figure 7.2.
It may be observed that in the Toposys 400 and 800 data, the histogram maximum is near the
tree top, but in the TopEye 200, 300 and 550 data sets, pulses have also been reflected from
the lower branches (see Figure 7.1).
Figure 7.2. Histograms of z co-ordinates of tree no. 28. In vertical axes are the number of samples. Top row:
TopEye 200 and TopEye 300, bottom row from left to right: Topeye 550, Toposys 400 and Toposys 800.
65
The point density and lack of pulses reflected from low branches had an effect on the crown
profiles derived from vector models. Tree crown profiles were not properly gained from
Toposys 400, Toposys 800 and Topeye 550 data sets, even if the average point density in
Toposys 400 was 20 points per square meter. The TopEye 300 and TopEye 200 data sets
resulted in the most complete crown profiles in vector models (Figures 6.15, 6.16 and 6.17).
7.2 Effect of overlapping flight lines
The data sets applied in the study differed from each other also in terms of flying geometry.
Toposys data sets were acquired by flying only one strip, whereas TopEye data sets were
collected from multiple strips. The latter method was expected to be better, since the tree is
scanned from more than one direction. The impact of overlapping strips was compensated by
performing new single tree pulse extraction using only one measurement strip for each tree in
TopEye 200 data. The vector model reconstruction was otherwise carried out in the same way
as the other data sets were processed previously and comparison with the tacheometer and
side view image measurements was repeated.
Table 7.2. Residuals of trunk location estimation.
Derivation of trunk location has been carried out
calculating average of crown hits.
Distance TopEye 200, points from two strips
TopEye 200, points from
one strip std (m) 0.74 0.81 bias (m) 1.18 1.37
Table 7.3. Residuals of trunk location estimation.
Derivation of trunk location has been carried out
using plane co-ordinates of the highest
measurement point.
Distance TopEye 200, points from two strips
TopEye 200, points from
one strip std (m) 0.79 0.81
Bias (m) 1.18 1.17
The results (see Table 7.2) demonstrated the hypothesis that trunk location estimation was
smaller accurate for trees that have been scanned only from one strip using average of crown
hits. Both biases and standard deviations increased in this case. The impact on trunk locations,
which were acquired applying plane co-ordinates of highest measurement point was found to
be less important, since accuracy did not decrease (Table 7.3). The estimated trunk location
was changed only if highest hit was measured from the reduced flight line.
66
7.3 Beam size and biases
The impact of beam size on the modelling accuracy was carried out by reducing data from the
Toposys 400 data set to the same point density as the Toposys 800 data set. As these materials
were acquired with same instrument at the same time, the only difference in the data is the
size of the beam ellipse, which was 40 cm in the data collected at 400 m and 80 cm in the data
collected at 800 m. It was assumed that biases with large beam size would have been less
concerning tree height, since the probability that pulses are reflected just from the crown top
would be higher with larger beams. However, even though biases increased with estimation
data, they were still smaller than in results derived with Toposys 800 data (Table 7.4).
Table 7.4. Results of tree height comparison with original Toposys 400 data, Toposys estimated data and
original Toposys 800 data.
Tree height Toposys 400 original Toposys 400 sparse Toposys 800
Std (m) 1.04 1.04 1.09 Bias (m) -0.24 -0.37 -0.77
The results were not as expected. In Yu & al (2004) it was suggested that a large beam size
improved the probability of laser pulse reflection from the crown top. However, the reduction
of laser points carried out in this thesis did not produce a similar scanning pattern as that
measured from a higher flight altitude. This is due to the uneven distance in scanning and
across the scan direction. Every second measurement point in the scanning order was reduced
in order to produce a pattern in which the point distance in flying direction would stay the
same as before the reduction, but the distance in the across–scan direction would double.
Unfortunately the number of sent pulses was not the same as the number of received pulses,
which were stored in the data, since some echoes were missed by the receiver. Therefore the
scanning pattern produced by reduction was more scattered than the pattern measured from a
higher altitude.
67
8 SUMMARY AND CONCLUSIONS
8.1 Summary and conclusions
The objective of this study was to provide a method for producing vector models of single
trees from laser scanner data in order to derive single tree geometry features. The method
included the following stages: point classification, digital terrain model production, extraction
of points from each tree, and vector model creation. Features extracted from laser-derived
models were tree height, crown height, trunk location, maximum crown width and crown
profile. Tree height and crown height represent the vertical geometry of a tree and trunk
location, maximum crown width, and crown profile the horizontal geometry. Since tree height
extraction requires information about the ground altitude at trunk locations, the digital terrain
model of the area was applied.
Also another method was introduced in the study, namely the side view imaging of trees for
collecting reference material for the laser-derived feature analysis. Trees in the test area were
imaged with predefined viewing geometry from two directions and the images were rectified
to a plane that is parallel to the tree trunk and perpendicular to the image capture direction.
The side view imaging method provided material that could not have been collected with
traditional forest inventory equipment. In addition to the crown and tree height, also side
profiles and the maximum width of the crown in the imaging direction were measured from
the images. If more than one image capture direction was applied, also trunk tilt could be
determined.
In order to test and evaluate the introduced methods, an experiment was carried out in the
Otaniemi study area applying five different laser data sets to reconstruct 50 trees in the area.
The features derived from laser models were compared to reference materials, which
consisted of tacheometer measurements, hypsometer measurements, and measurements
acquired from side view images.
According to the experiment, this method provided vector models of trees with the following
parameters and their accuracies: In vertical direction, tree heights were derived with a
negative bias of approximately 0.5 m and a standard deviation of 1 m. The most accurate
results were gained using TopEye 300 data, where the bias was only –0.06 m and STD was
68
1.02 m. However, crown heights were derived with only moderate accuracy; the negative
biases varied from -0.84 m to -3.84 m and standard deviations from 2.98 m to 5.38 m. These
comparisons were carried out by applying side view image measurements, since their
accuracy was found to be better than the accuracy of hypsometer measurements.
In the horizontal direction, the trunk locations were estimated from the models with a bias of
0.74 to 0.84 m and an STD of 1.18 m to 1.44 m when single tree point extraction was
performed manually. These results were obtained deriving trunk location as average of crown
class points. The other approach, the segmentation-based method, provided similar results,
except for biases that were slightly smaller, from 0.58 m to 0.74 m. The crown widths were
gained with negative biases from –0.78 to 0.46 m and STD from 1.38 to 2.48 m, which is a
moderate accuracy. The result was confirmed in a visual analysis of crown profiles compared
to the profiles derived from side view images.
The DTM accuracy was also considered, since errors in DTM emphasise errors in tree height
derivation. The experiences indicated that ground cover type and openness of the area
affected the laser-derived DTM accuracy, and that DTM is typically located slightly above the
ground truth. The worst standard deviation of the residuals was 15 cm, but the best and more
common was 7 cm. In the comparison, the DTM accuracy was highest on asphalt-paved roads
and decreased in tree-shadowed undervegetation areas.
All results indicated that tree crown shape is underestimated in the laser models. The tree and
crown heights and crown widths were negatively biased, and these results were gained with
all five different laser data sets. The reasons for underestimation were suggested to be (1)
pulse distribution of laser measurement and (2) shape of the pulse signal reflected from
vegetation.
A tree has a complicated shape. Branches grow unsymmetrically around the trunk and have
various lengths, and the trunk itself may be in an oblique position. In laser scanning, pulses
are not directed to a certain part of the measured object, but distributed in the area by a
scanning mechanism. Therefore they do not necessarily reflect from the desired part of the
surface, such as tips of the branches or top of the tree. As no data simulation is included in the
method, the tree parts that have not reflected laser pulses are not presented in the models. In
69
the visual analysis of laser-derived crown profiles versus image-derived profiles, it was
observed that crown parts with only a few laser hits were incomplete in the models. The
tendency is the same in the vertical dimension; high negative biases in crown height
estimation indicate that laser pulses have not reflected from the lower branches of the tree
crown, and therefore they are not presented in the models.
The visual analysis of return signals originated from vegetation revealed the uncertainty of
ranging measurement, since the shape of the return signal is long and varying instead of one
clear peak. When applying different algorithms, the signal could be determined as returned in
several positions and measurement vector having several lengths. Underestimation of laser-
derived tree model was interpreted so that the crown surface, which is the first to reflect the
pulses, is not presented by laser measurements, and the strongest reflection has originated
from within the crown instead of from the surface.
The laser scanning parameters affected the vector model application method. High pulse
density improves tree reconstruction and derivation of 3D geometry features. If tree height is
the only parameter to be derived, a steep scanning angle is more suitable, since shadow areas
behind impenetrable objects are small. In steep scanning, pulses are also able to reach the
ground, and these pulses are suitable for digital terrain model production. However, oblique
scanning is better than steep scanning in order to gain pulses from lower branches of the tree
in open areas, where pulses are able to reach the branches without interfering with
neighbouring tree crowns. To estimate trunk location, trees should be scanned from more than
one direction in order to make sure that both sides of the crown reflect pulses.
8.2 Discussion
The laser method introduced here provides a way to extract features of single trees from laser
scanner data. The application area of the method is individual treewise forest inventory. If a
DTM is produced from laser data, no other external information is required.
However, the method is not fully automatic. In single tree extraction, two approaches were
applied to delineate tree crowns, a manual and a segmentation-based approach. The
segmentation-based method represents a fully automatic approach, whereas manual
delineation was carried out by a human interpreting the CHM surface. As previous studies
70
have shown, segmentation is successful only if there is enough space between a tree and its
neighbours. In a dense forest, where tree crowns are joined, a single crown is either split to
several segments or merged with surrounding trees. In this study, the proportion of correctly
detected trees varied from 34 % to 64 % depending on the data set. Even with the best
results, the amount is so low that the usability of segmentation-based approach is
questionable. Therefore, manual delineation is recommended even if it is time-consuming and
expensive.
The other application area of the laser method is 3D visualisation of forest areas. Typically
visualisation is carried out for the purpose of 3D management and planning in the forest areas.
An example of one area is illustrated in Figure 8.1. Since high accuracy is not required in
visualisation, segmentation may take place and automate the whole process.
Figure 8.1. Kalkkinen test area processed with introduced laser method (Pyysalo, 1999).
8.3 Future work
This thesis was limited to the topic of single tree parameters that are available from a 3D
vector model. Thus, parameters such as timber volume or diameter at breast height were not
considered, since laser beams only rarely reflect straight from the trunk. These parameters,
however, have great economic importance in forest inventory. The next step in the tree
parameters evaluation would be to use the derived parameters as input values in timber
volume estimation models.
71
Another approach to derived single tree features would be to match 3D primitives of single
trees to a point cloud after the delineation of tree crowns. A priori knowledge of tree species
could be used, as well as information about the tree height distribution in the area. In this
study it was found that the lower parts of the tree crown are only moderately presented in the
data set. Also when measurement is carried out by applying oblique scanning, the opposite
side of the crown is shadowed by the side of the crown facing the scanner and therefore
unevenly presented in the data. The matching of trunk-symmetric 3D models would estimate
the areas where measurement information is limited.
Laser scanning implementation has developed during the preparation of this thesis. The first
data was collected with an instrument capable of registering only one echo, first or last. This
system was followed by an instrument capable of registering two echoes, first and last, and
then by an instrument registering even five echoes. The possibility of registering intensity was
also included in the system and measurement frequency has increased all the time. The latest
step in development has been the full waveform measurement mode, which has also been
discussed in this thesis. Ranging frequency with the present instrument is already high, but the
accuracy of single measurement is only moderate. The full waveform measurement provides
an opportunity to improve single point accuracy and estimate the single point quality. The
author assumes that future development concerning laser scanning applicability in tree
modelling will take into account quality-weighted laser point processing and signal analysis.
72
REFERENCES
Ahokas, E., Kaartinen, H., Matikainen, L., Hyyppä, J., Hyyppä, H., 2002. Accuracy of
high-pulse-rate laser scanners for digital target models. In: Observing our environment from
space. New solutions for a new millennium. Proceedings of the 21st EARSeL Symposium,
Paris, 14-16 May, 2001. Balkema Publishers 2002, pp. 175-178.
Aldred, A.H. and Bonnor, G.M., 1985. Application of airborne lasers to forest surveys.
Information Report PI-X-51, Canadian Forestry Service, Petawawa national Forestry Institute,
62 p.
Andersen, H-E, Reutebuch, S., Schreuder, G., 2002. Bayesian Object Recognition for the
Analysis of Complex Forest Scenes in Airborne Laser Scanner Data. International Archives of
the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol 34, part 3A, pp-
35-41.
Batsavias, E.P., 1999. Airborne laser scanning: basic relations and formulas, ISPRS Journal of
Photogrammetry and Remote Sensing 53, pp.199-214
Brandtberg, T., Warner T., Landenberger, R., McGraw, J., 2003. Detection and analysis of
individual leaf-off tree crowns in small footprint, high sampling density lidar data from the
eastern deciduous forest in North America. Remote Sensing of Environment 85, pp. 290-303.
Burman, H, 2000. Calibration and Orientation of Airborne Image and Laser Scanner Data
Using GPS and INS. Dissertation, Kungl Tekniska Högskolan.
Elmqvist, M., Jungert, E., Lantz, F., Persson, Å., Söderman, U., 2001. Terrain modelling
and analysis using laser scanner data. estimation of laser radar data using active shape models.
International Archives of Photogrammetry and Remote Sensing. Vol. 34-3/W4, pp. 219-227.
Friedlaender, H., Koch, B., 2000. First experience in the application of laserscanner data for
the assessment of vertical and horizontal forest structures. In: International Archives of
Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. pp 693-700.
73
Gaveau, D., Hill, R., 2003. Quantifying canopy height underestimation by laser pulse
penetration in small-footprint airborne laser scanning data. Canadian Journal of Remote Sensing
29, pp. 650–657.
Gougeon, F., Leckie, D., 2003. Forest information extraction from high spatial resolution
images using an individual tree crown approach. Information report BC-X-396, Natural
Resources Canada, Canadian Forest Service, Pacific Forestry Centre, 26 p.
Gougeon, F., Moore, T., Classification indiovidualle des arbres á partir d’images á haute
resolution spatiale. Télédétection et gestion des resources Vol. VI, pp. 185-196. (In French)
Haggren, H., Manninen, T., Peräläinen, I., Pesonen, J., Pöntinen, P,; Rantasuo, M., 1995.
Airborne 3D-profilometer. VTT Research Notes : 1667. 19 pages.
Hyyppä J., Pyysalo U., Hyyppä H., Haggren H., Ruppert G., 2000. Accuracy of laser
scanning for DTM generation in forested areas. Proceedings of SPIE. Vol. 4035, pp. 119-130.
Hyyppä, J. and Inkinen, M., 1999. Detecting and estimating attributes for single trees using
laser scanner. The Photogrammetric Journal of Finland, 16, pp. 27-42.
Hyyppä, J., Schardt, M., Haggrén, H., Koch, B., Lohr, U., Scherrer, H.U., Paananen, R.,
Luukkonen, H., Ziegler, M., Hyyppä, H., Pyysalo, U., Friedländer, H., Uuttera, J.,
Wagner, S., Inkinen, M., Wimmer, A., Kukko, A., Ahokas, E., Karjalainen, M., 2001.
HIGH-SCAN: The first European-wide attempt to derive single-tree information from
laserscanner data, The Photogrammetric Journal of Finland 17, pp. 58-68.
JHS 154 ETRS89 -järjestelmään liittyvät karttaprojektiot, tasokoordinaatistot ja karttalehtijako,
Julkisen hallinnon suositukset nro 154.
Kraus, K., Pfeifer, N., 1998. Determination of terrain models in wooded areas with airborne
laser scanner data. ISPRS Journal of Photogrammetry and Remote Sensing 53, pp.193-203.
74
Leckie, D., Gougeon, F., Hill, D.,Quinn, R., Armstrong, L., Shreenan, R., 2003. Combined
high-density lidar and multispectral imagery for individual tree crown analysis. Canadian
Journal of Remote Sensing 29, No. 5, pp. 633–649.
Lefsky, M., Cohen, W., Acker, S., Parker, G., Spies, T., Harding, D., 1999. Lidar remote
sensing of the canopy structure and biophysical properties of Douglas-fir western hemlock
forests. Remote Sensing of Environment 70, pp. 339-361.
Magnussen, S., Boudewyn, P., 1998. Derivations of stand heights from airborne laser scanner
data with canopy-based quantile estimators. Canadian Journal of Forest Research 28, pp 1016-
1031.
Magnussen, S., Eggermont, P., LaRiccia, V.N., 1999. Recovering tree heights from airborne
laser scanner data. Forest Science 45, pp. 407-422.
Maltamo, M., Mustonen, K., Hyyppä, J., Pitkänen, J. and Yu, X. 2004. The accuracy of
estimating individual tree variables with airborne laser scanning in a boreal nature reserve.
Canadian Journal of Forest Research 34: 1791-1801.
Means, J., Acker, S., Harding, D., Blair, J., Lefsky, M., Cohen, W., Harmon, M., McKee,
A., 1999. Use of large-footprint scanning airborne lidar to estimate forest stand characteristics
in the western cascades of Oregon. Remote Sensing of Environment 67, pp. 298-308.
Mikhail E. M., Bethel J. S., and McGlone, 2001 Introduction to Modern Photogrammetry.
John Wiley and Sons. New York. ISBN 0-471-30924-9, pp. 81-106
Morsdorf, F., Meier, E., Allgöwer, B., Nüesch, D., 2003. Clustering in airborne laser scanning
raw data for segmentation of single trees. International Archives of the Photogrammetry,
Remote Sensing and Spatial Information Sciences, Vol 34, part 3/W13, pp. 27-33.
Næsset, E. ,Økland, T., 2002. Estimating tree height and tree crown properties using airborne
scanning laser in a boreal nature reserve. Remote Sensing of Environment 79, pp. 105-115.
75
Næsset, E. 1997a. Determination of mean tree height of forest stands using airborne laser
scanner data. ISPRS Journal of Photogrammetry and Remote Sensing 52, pp. 49-56.
Næsset, E. 1997b. Estimating timber volume of forest stands using airborne laser scanner data.
Remote Sensing of Environment 61, pp-246-253.
Næsset, E., 2002. Predicting forest stand characteristics with airborne scanning laser using a
practical two-stage procedure and field data. Remote Sensing of Environment 80, pp.88-99.
Nelson, R., Krabill, W. and Maclean, G., 1984. Determining forest canopy characteristics
using airborne laser data, Remote Sensing of Environment 15, pp. 201-212.
Ollikainen, M. 1993. GPS-koordinaattien muuntaminen Kartastokoordinaateiksi. Geodeettisen
laitoksen tiedonantoja, No. 8. 31 pages, Helsinki..
Persson., Å, Holmgren, J., Söderman, U., 2002. Detecting and measuring individual trees
using an airborne laser scanner. Photogrammetric Engineering and Remote Sensing 68, No 9,
pp. 925-932.
Popescu, S., Wynne, R., Nelson, R., 2003. Measuring individual tree crown diameter with
lidar and assessing its influence on estimating forest volume and biomass. Canadian Journal of
Remote Sensing. 29, No. 5, pp. 564–577.
Pyysalo, U. A Method to create a three-dimensional forest model from laser scanner data, The
Photogrammetric Journal of Finland 17, No. 1. 2000, pp. 34-42
Pyysalo, U., 2000. Generation of elevation models in wooded areas from a three dimensional
point cloud measured by laser scanning. MSc Thesis, Helsinki University of Technology,
Espoo, Finland, 68 p.
Pyysalo, U., Hyyppä, H. Reconstructing Tree Crowns from Laser Scanner Data for Feature
Extraction. International Society for Photogrammetry and Remote Sensing - ISPRS
Commission III Symposium (PCV'02). September 9-13, 2002. Graz. Austria.
76
Reutebuch, S., McGaughey, R., Andersen, H., Carson, W., 2003. Accuracy of a high-
resolution lidar terrain model under a conifer forest canopy. Canadian Journal of Remote
Sensing, 29, pp. 527-535.
Rönnholm, P., J. Hyyppä, H. Hyyppä, H. Haggrén, X. Yu, and H. Kaartinen, 2004.
Calibration of Laser-derived Tree Height Estimates by means of Photogrammetric Techniques,
Scanfinavian Journal of Forest Research, Vol. 19, No. 6, pp. 524-528.
Rönnholm, P., 2005. Registration and Fusion of Terrestrial Digital Images and Airborne Laser
Point Clouds. Licentiate’s thesis. Helsinki University of Technology. 74 p.
Sithole, G., 2001. Filtering of laser altimetry data using a slope adaptive filter. International
Archives of Photogrammetry and Remote Sensing. Vol. 34-3/W4, pp. 203-210.
Solodukhin, V., Zukov, A., Mazugin, I., 1977. Possibilities of laser aerial photography for
forest profiling, Lesnoe Khozyaisto (Forest Management) 10, pp. 53-58 (in Russian).
Straub. B., 2003. A top-down operator for the automatic extraction of trees - concept and
performance evaluation. Proceedings of the ISPRS working group III/3 workshop `3-D
reconstruction from airborne laserscanner and InSAR data' Dresden, Germany 8-10 October
2003, pp. 34-39.
Takeda, H., 2004. Ground surface estimation in dense forest. The International Archives of
Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol. 35, part B3, pp. 1016-
1023.
Wack, R. Wimmer, A., 2002. Digital terrain models from airborne laser scanner data – a grid
based approach. . International Archives of Photogrammetry and Remote Sensing. Vol. 35, Part
3B, pp. 293-296.
77
Wack, R., Schardt. Lohr, U., Barrucho, L., Oliveira, T., 2003. Forest inventory for
Eucalyptus plantations based on airborne laser scanner data. International Archives of the
Photogrammetry, Remote Sensing and Spatial Information Aciences, Vol 34, part 3/W13, pp.
40-46.
Wehr, A., Lohr, U., 1999. Airborne laser scanning - an introduction and overview. ISPRS
Journal of Photogrammetry and Remote Sensing Vol. 54, pp. 68-82.
Vosselman and Maas, 2001. Adjustment and filtering of raw laser altimetry data. Proceedings
of OEEPE workshop on airborne laserscanning and interferometric SAR for detailed digital
elevation models, Royal Institute of Technology, Stockholm, Sweden, 11 p.
Yu, X., Hyyppä, J., Kaartinen, H., Maltamo, M., 2004. Automatic detection of harvested
trees and determination of forest growth using airborne laser scanning. Remote Sensing of
Environment 90, pp. 451-462.
Yu, X., Hyyppä, J., Hyyppä, H. and M. Maltamo, 2004. Effects of flight altitude on tree
height estimation using airborne laser scanning. International Archives of Photogrammetry,
Remote Sensing and Spatial Information Sciences, XXXVI (8/W2), pp. 96-101.
http://www.leica-geosystems.com
78