Species discrimination and individual tree detection for [608258]

Elsevier Editorial System(tm) for Science of
the Total Environment
Manuscript Draft

Manuscript Number:

Title: Species discrimination and individual tree detection for
predicting main dendrometric characteristics in mixed temperate forests
by use of Airborne Laser Scanning and ultra -high-resolution imagery

Article Type: VSI:INCDS CONFERENCE 85

Keywords: ALS, UAV, OBIA, forest inventory, Monte Carlo simulation

Corresponding Au thor: Dr. Bogdan Apostol, Ph.D.

Corresponding Author's Institution: National Institute for Research and
Development in Forestry (INCDS) ,,Marin Dracea"

First Author: Bogdan Apostol, Ph.D.

Order of Authors: Bogdan Apostol, Ph.D.; Marius Petrila, Ph .D.; Adrian
Lorent; Albert Ciceu; Vladimir Gancz, Ph.D.; Ovidiu Badea, Ph.D.

Abstract: This study aims to investigate the combined use of two types of
remote sensing data — ALS derived and digital aerial photogrammetry data
(based on imagery collected by airborne UAV sensors) — along with
intensive field measurements for extracting and predicting tree and stand
parameters in even -aged mixed forests. The study is located in South West
Romania and analyzes data collected from mixed -species plots. The main
tree species within each plot are Norway spruce (Picea abies L. Karst.)
and Beech (Fagus sylvatica L.).
The ALS data were used to extract the digital terrain model (DTM),
digital surface model (DSM) and normalized canopy height model (CHM).
Object-Based Image Analysis (OBIA) classification was performed to
automatically detect and separate the main tree species. A local
filtering algorithm with a canopy -height based variable window size was
applied to identify the position, height and crown diameter of th e main
tree species within each plot. The filter was separately applied for each
of the plots and for the areas covered with Norway spruce and beech
trees, respectively (i.e. as resulted from OBIA classification).
The dbh was predicted based on ALS data by statistical Monte Carlo
simulations and a linear regression model that relates field dbh for each
tree species with their corresponding ALS -derived tree height and crown
diameter. The overall RMSE for each of the tree species within all the
plots was 5. 8 cm for the Norway spruce trees, respectively 5.9 cm for the
beech trees. The results indicate a higher individual tree detection rate
and subsequently a more precise estimation of dendrometric parameters for
Norway spruce compared to beech trees located in spruce -beech even -aged
mixed stands. Further investigations are required, particularly in the
case of choosing the best method for individual tree detection of beech
trees located in temperate even -aged mixed stands.

Suggested Reviewers: Iosif Vorove ncii Ph.D.
Professor, “TRANSILVANIA” University of Brașov, Romania
[anonimizat]
research experience in remote sensing and forestry field

Mihai-Daniel Niță Ph.D.
“TRANSILVANIA” University of Brașov, Romania
[anonimizat]
research experience in r emote sensing and forestry field

Thomas Schneider Ph.D.
Technical University of München, Germany
[anonimizat]
research experience in remote sensing and forestry field

Martin Lorenz Ph.D.
Thünen-Institut für Internationale Waldwirtschaft und Fors tökonomie
Arbeitsbereich Wald und Gesellschaft
[anonimizat]
research experience in forest management, remote sensing forestry
applications

Curtis Woodcock Ph.D.
Professor, Boston University
curtis@bu.edu

Opposed Reviewers:

To:
Editor -in-Chief of Science of the Total
Environment
Place, date
Bucharest, 14.05.2019
Cover letter
Dear Editor -in-Chief,

We would like to submit our manuscript entitled “ Species discrimination and individual tree
detection for predicting main dendrom etric characteristics in mixed temperate forests by use of
Airborne Laser Scanning and ultra -high -resolution imagery ”, authors: Bogdan Apostol1,*, Marius
Petrila1, Adrian Lorenț1,2, Albert Ciceu1, Vladimir Gancz1, Ovidiu Badea1,2
1National Institute for Research and Development in Forestry (INCDS),“Marin Drăcea”, Romania
2Faculty of Silviculture and Forest Engineering, “Transilvania” University of Brașov, Romania

In order to ensure an efficient and sustainable management of the forest ecosystems, determined
by the environmental conditions and human economic interests, the combined use of two types of
remote sensing data — ALS derived and digi tal aerial photogrammet ry data can bring significant
improvements for forest assessment . This issue was well studied for boreal forests, but for precision
forestry in mixed forests there is a need for better tools for quantitative and qualitative assessment
of trees and forest s tands parameters . Hopefully our paper is a small step further in creating these
improved methodologies.

Sincerely,
On behalf of all authors
Bogdan Apostol

*Corresponding author at: National Institute for Research and Development in Forestry (INCDS) “Marin
Drăcea” , Bdul Eroilor 128, Voluntari, 077190, Ilfov, Romania , e-mail address: bogdanap_ro@yahoo.com Cover Letter

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65 Title: Species discrimination and individual tree detection for predicting main
dendrometric characteristics in mixed temperate forests by use of Airborne Laser
Scanning and ultra -high -resolution imagery : Bogdan Apostol1, Marius Petrila1, Adrian
Lorenț1,2, Albert Ciceu1, Vladimir Gancz1, Ovidiu Badea1,2
Affiliations:
1National Institute for Research and De velopment in Forestry (INCDS),“Marin Drăcea”,
Romania
2Faculty of Silviculture and Forest Engineering, “Transilvania” University of Brașov, Romania
bogdanap_ro@yahoo.com
marius.petrila@icas.ro
adrian.lorent@icas.ro
albert.ciceu@ica s.ro
vladimir.ga ncz@icas.ro
obadea@icas.ro

Corresponding author Bogdan Apostol , bogdanap_ro@yahoo.com ,

*Title Page

*Graphical Abstract

 OBIA classification accuracy of the tree species was between 73.9 -77.3%
 Individual tree detection rate was higher for the spruce than for the beech trees
 The tree dbh was predicted based on ALS data by statistical Monte Carlo simulations
 No statistically s ignificant difference was revealed between the ALS -predicted and
field-measured dbh.
 The volume of the ALS -detected spruce trees represents up to 79% from the total
spruce trees field volum e
*Highlights (for review : 3 to 5 bullet points (maximum 85 characters including spaces per bullet point)

1
Species discrimination and individual tree detection for predicting main dendrometric 1
characteristics in mixed temperate forests by use of Airborne Laser Scanning and ultra – 2
high -resolution imagery 3
4
Bogdan Apostol1,*, Marius Petrila1, Adrian Lorenț1,2, Albe rt Ciceu1, Vladimir Gancz1, Ovidiu 5
Badea1,2 6
1National Institute for Research and Deve lopment in Forestry (INCDS) ,“Marin Drăcea”, 7
Romania 8
2Faculty of Silviculture and Forest Engineering, “Transilvania” University of Brașov, 9
Romania 10
11
12
Abstract: This study ai ms to investigate the combined use of two types of remote sensing 13
data — ALS derived and digital aerial photogrammetry data (based on imagery collected by 14
airborne UAV sensors) — along with intensive field measurements for extracting and 15
predi cting tree an d stand parameters in even -aged mixed forests. The study is located in 16
South West Romania and analyzes data collected from mixed -species plots. The main tree 17
species within each plot are Norway spruce ( Picea abies L. Karst.) and Beech ( Fagus 18
sylvatica L.). 19
The ALS data were used to extract the digital terrain model (DTM), digital surface model 20
(DSM) and normalized canopy height model (CHM). Object -Based Image Analysis (OBIA) 21
classification was performed to automatically detect and separate the main tree spe cies. A 22
local filtering algorithm with a canopy -height based variable window size was applied to 23
identify the position, height and crown diameter of the main tree species within each plot. 24
The filter was separately applied for each of the plot s and for the areas covered with Norway 25
spruce and beech trees, respectively (i.e. as resulted from OBIA classification). 26

*Corresponding author at: National Institute for Research and Development in Forestry (INCDS) “Marin
Drăcea” , Bdul Eroilor 128, Voluntari, 077190, Ilfov, Romania , e-mail address: bogdanap_ro@yahoo.com *Manuscript (double-spaced and continuously LINE and PAGE numbered)-for final publication
Click here to view linked References

2
The dbh was predicted based on ALS data by statistical Monte Carlo simulations and a linear 27
regression model that relates field dbh f or each tree species with their corresponding ALS – 28
derived tree height and crown diameter. The overall RMSE for each of the tree species within 29
all the plots was 5.8 cm for the Norway spruce trees, respectively 5.9 cm for the beech trees. 30
The results indica te a higher i ndividual tree detection rate and subsequently a more precise 31
estimation of dendrometric parameters for Norway spruce compared to beech trees located in 32
spruce -beech even -aged mixed stands. Further investigations are required, particularly in the 33
case of c hoosing the best method for individual tree detection of beech trees located in 34
temperate even -aged mixed stands. 35
Keywords: ALS , UAV, OBIA, forest inventory , Monte Carlo simulation 36
37
1. Introduction 38
39
Recently developed remote sensing technologies offer the po ssibility of a detailed tri – 40
dimensional representation of the topography and spatial positioning of trees and thus allow 41
for more accurate modelling of forests by measuring the dendrometric variables of trees and 42
stands, but also for qualitati ve characteri stics such as identifying species and estimating 43
composition. 44
With now more than 30 years’ research, continual improvement as well as increasing 45
accessibility of the technology, airborne laser scanning data (ALS) has been able to gain 46
maturit y in the fiel d of forest inventory and is now moving towards operational applications, 47
mainly in the Scandinavian countries (Hyyppä et al. 2012; White et al. 2013). 48
Forest parameter retrieval methods from ALS data can be categorized under two major 49
approaches: individu al tree detection and the area -based approach (i.e based on statistical 50
canopy height distribution), respectively (Hyyppä et al. 2008). The methods for individual 51

3
tree extraction information from ALS can also be classified into three ca tegories (Duncanson 52
et al. 2014): local maxima detection and expansion, watershed -based delineation and point – 53
cloud clustering. Local maxima approaches rely on a rasterized canopy height model (CHM) 54
computed generally from the first returns of ALS or the m aximum height values of all returns 55
so that the algorithm fails to intercept the understory (Hyyppä et al. 2001; Solberg et al. 2006; 56
Wang et al. 2016), as well as having issues related to segmenting neighbouring trees 57
(Kaartinen et al. 2012; Reitberger et al. 2009) when the CHM reconstructs the outer part of 58
the crowns. However, the tree detection rate and parameter retrieval (e.g. heights, volumes or 59
biomass) is generally satisfactory in the case of boreal forests (Apostol et al. 2016; Hyyppä et 60
al. 2001; Persson et al. 2002 ; Popescu, 2007; Yu et al. 2011). Watershed segmentation 61
determines individual tree crowns using a CHM by segmenting the inverted raster canopy 62
surface into equivalent individual hydrologic drainage basins (Edson and Wing, 2011). While 63
some authors report certain advantages with the watershed segmentation (e.g. better 64
identifying understory and overlapping trees) (Duncanson et al. 2014; Edson and Wing, 65
2011), others have argued that using a downwards local maxima region growing algorithm 66
ensures a better sh ape control and more appropriate crown -like segments (Solberg et al. 67
2006). Finally, cloud clustering or the point -based method uses the row 3D point cloud which 68
takes advantage of using all the information stored in the laser data. The method has gained 69
significant attention recently, particularly because the computer infrastructure is becoming 70
sufficiently powerful to handle the significant amounts of information contained in the ALS 71
point cloud; thus, this approach may yield good pros pects for deriving f orest metrics (Ferraz 72
et al. 2012, Tiede et al. 2012). Finding a flexible algorithm for ALS data that performs well 73
across varying forest structures is still an open issue because challenges have been raised 74
mostly in the case of dense, multi -layered, even -aged, mixed (broadleaves and coniferous) 75
forests. 76

4
Individual tree detection and delineation between coniferous species (usually pine and 77
spruce) are easier in boreal forests, where crowns are significantly more clearly delimited 78
than i n mixed forests, whe re biomass studies also occasionally use trees of known species 79
(Kankare et al., 2013) . 80
Kaartinen et al. (2012) conducted an extensive experiment regarding how the methods of 81
automatic individual tree extraction parameters (i.e. location, detection rate, tree heights and 82
crown width) affect the accuracy of the results, for a mixed (resinous an d broadl eaves) 83
heterogeneous forest using discrete (first -last echoes) ALS data with point densities varying 84
from 2 to 8 points per m2. A total of 13 automated methods were analyzed for the same ALS 85
data by an international consortium consisting of ten par tners fr om eight countries, with each 86
partner employing its own strategy for tree detection and crown segmentation. The results of 87
the different approaches were quite heterogenous because the percentage of detected trees 88
varied between 25 to 102%, along wi th the r esults for crown area appearing to vary 89
significantly between the models. A more recent international benchmarking on 3 -D canopy 90
structure modelling based on ALS was reported by Wang et al. (2016). Five individual tree 91
detection methods were tested by five partners from four countries, where two were raster 92
based, two were hybrid and one was point based. They found that canopy structure plays a 93
considerable role in the detection accuracy and would even have a higher impact than the tree 94
species or s pecies c omposition in a stand. They recommended use of the point cloud data for 95
the detection of intermediate and suppressed trees stating that the point density would 96
significantly influence the performance of point -based methods. 97
One study compared senso rs mount ed on aircraft (ALS data versus aerial hyperspectral 98
imagery) for species and tree delineation, using manual and automatic approaches, but the 99
broadleaves species was always poorly classified than coniferous (Dalponte et al., 2014) . In 100
another study, fusion between ALS dat a and ae rial imagery, along with field sample plots or 101

5
3D manual species delineation on imagery as reference data, was used for species 102
discrimination (Heinzel et al., 2008) . 103
ALS technology is still subject to several drawbacks such as the lack of spectral information 104
(which makes it not particularly suitable for species recognition) and the still relatively high 105
acquisition costs that limit i ts usage, particularly in the case of multitemporal analysis. The 106
requirement for better results has led other researchers to combine ALS data with other data 107
sources from satellite or aerial sensors (Räsänen et al. 2014 , Wittke et al. 2019) . 108
One n ew promi sing emerging technology seems to fill these gaps in the form of ultra -high- 109
resolution aerial images captured with digital (often multispectral) cameras, installed on 110
unmanned aerial vehicles (UAV s), which can be a more effective alternative to con ventiona l 111
optical sources such as satellite and aerial platforms. The new improvements in digital 112
photogrammetry also allow computer vision processing of aerial imagery by Structure from 113
Motion algorithms (SfM). UAVs can combine high spatial resolution and rapid revisit times 114
together with lower operational costs and complexity (Chianucci et al. 2016). It also offers an 115
economic alternative data source that supports the retrieval of tree dimensionality and 116
location information (Gatziolis et al. 2015). One case used only UAV -based photogrammetric 117
3D point clouds obtained from RGB and hyperspectral sensors for individual tree detection 118
and tree species classification in boreal forests, with a d etection accuracy of between 40% 119
and 90%, depending on the area characteristics (Nevalainen et al. 2017) . 120
To the authors’ knowledge, no study has evaluated the potential of ALS -UAV data to model 121
forest biophysical properties in mixed forests. The aim of the study is to investigate the 122
combine d use of both types of remote sensing data — ALS derived and digital aerial 123
photogrammetry data (based on imagery collected by airborne UAV sensors) — along with 124
intensive field measurements for extracting tree and stand parameters in even -aged mix ed 125
fores ts by statistical Monte Carlo simulations. The first objective was the extraction of the 126

6
spruce and beech crowns, and the canopy gaps based on Object -Based Image Analysis 127
(OBIA) of the UAV data. A second objective was the individual tree detection and the 128
prediction of tree dbh based on ALS and UAV data. The ALS -predicted dbh and the derived 129
tree height were used for tree volume assessment. 130
131
2. Material and methods 132
2.1 Study area 133
The study was conducted in four square plots with an area of 1 ha (100 X 100 m) each, within 134
even -aged mixed forest stands of Norway spruce (Picea abies L. Karst.), common beech 135
(Fagus sylvatica L.) and other hardwood tree species, all located in South West Romania, in 136
the Southern Carpathians (Experimental Forest District Caranse beș) (Figure 1). The current 137
forest management plans of the Experimental Forest District Caransebeș describes the stands 138
as 80 and 85 years old, respectively. 139
140
Fig. 1. Location of the study . 141
2.2 Field measurements 142
The square plots (41, 44, 38 and 40) were pl aced using Field Map equipment (Vopěnka and 143
Černý, 2006). The position and crown projections of the trees were measured in a local 144
coordinate system using the Field Map system. At the same time, descriptive data such as the 145
vitality and competitive intensi ty of the trees, referred to as Kraft classification, were visually 146

7
determined in the field and introduced in the Field Map database. There were five social 147
classes assigned (Kraft, 1884) which were: predominant (class I), dominant (class II), co – 148
dominant (class III), dominated (class IV) and suppressed (class V). A Vertex IV 149
inclinometer was used to measure the tree heights, while measurements for tree dbh were 150
performed using an electronic caliper. Both measurements were then imported into the Field 151
Map D ata Collector software. The position of one corner of each plot was determined using a 152
Trimble Geo XH receiver equipped with a Zephir II external antenna. The corner coordinates 153
were used to convert the local coordinate system of the field measurements int o a national 154
projection system (Stereographic 1970 projection system). Sampling campaign campaigns 155
took place between 2015 and 2018; thus, each plot was inventoried in a different year: plot 41 156
was field measured in 2015, plot 44 in 2016, plot 38 in the 20 17 and plot 40 in 2018. 157
There were 2,069 trees measured within all the plots, with the descriptive statistics of the tree 158
species given in Table 1. 159
Table 1. Description of the forest characteristics with in all the plot s. 160
Plot Size
(ha) Tree
species Densi ty
(stems*ha-1) Height (m) Kraft Class
min max mean I II III IV V
41 1 Norway
spruce 208 5.2 41 29.22 57 61 48 18 24
Beech 183 4.7 39.5 25.10 16 57 57 49 4
Other
species 19 14.2 37.5 28.99 2 10 6 0 1
TOTAL 410 4.7 41 27.37 75 128 111 67 29
44 1 Norway
spruce 199 4 39.3 28.92 74 66 21 12 26
Beech 81 6.4 32.2 25.38 3 34 27 15 2
Other
species 48 7.7 37.4 23.55 4 7 20 16 1
TOTAL 328 4 39.3 27.26 81 107 68 43 29
38 1 Norway 201 4.4 35.6 25.44 86 44 28 12 31

8
spruce
Beech 394 4.9 34.3 22.33 34 180 133 37 10
Other
species 69 3.3 30.9 19.03 10 28 5 18 8
TOTAL 664 3.3 35.6 22.93 130 252 166 67 49
40 1 Norway
spruce 350 4 34.9 26.08 72 130 74 23 51
Beech 220 4.8 30.3 19.61 8 84 118 10
Other
species 97 4 31.8 18.71 1 10 31 39 16
TOT AL 667 4 34.9 22.87 73 148 189 180 77
161
2.3 Airborne Laser Scanning data 162
The ALS data were collected in July 2015 by a specialized company, using a DA42 MPP 163
aircraft. The plane was equipped with a full -wave Riegl LMS -Q780 laser scanner with a final 164
laser point density of 5 –8 points m -2. Details regarding the ALS data are given in Table 2. 165
The flight plan was designed to cover the forest of the Experimental Forest District 166
Caransebes. In total, 56 flight segments were flown for recording the ALS data, amounting to 167
263 kilometres. The ALS data were delivered unclassified and in th e Stereographic 1970 168
projection system. 169
Table 2. Characteristics of ALS data . 170
Parameter Name /Value
Sensor Airborne Lidar Scanner ‘digital full
waveform analysis ’, RIEGL Q780
Aircraft Diamond DA42 MPP aircraft
Mean f light altitude 808 m AGL
Field of view (FOV) >600
Pulse Repetition Rate (PRR): >400 kHz
Scanning speed > 266 ,000 measurements sec-1
Point density 5–8 points m-2
Altimetric precision 20 cm
Total planimetric precision 20 cm
171
2.4 Unmanned Aerial Vehicle (UAV) data 172

9
To discriminate between coniferous and deciduous trees from the plot, respectively Norway 173
spruce and be ech trees, aerial imagery was captured on 29 September 2016 using a SenseFly 174
eBee RTK fixed -wing mapping drone . The UAV is equipped with a LI/L2 global nav igation 175
satellite system (GNSS)/real time kinematic (RTK) receiver capable of receiving corrections 176
from either ground or virtual reference stations via its flight control software and ground 177
modem. According to the vendor specifications (SenseFly 2014), t his approach ensures a 178
relative orthomosaic/3D model accuracy of 1 –3 x GSD, and absolute horizont al/vertical 179
accuracy of down to 3/5 cm, without ground control points. 180
The UAV was equipped with a Canon S110 RGB frame camera for imagery acquisition. The 181
camera has a focal length of 5.32 mm and a resolution of 4000 x 3000 pixels and delivers 182
imagery in jpeg format in blue (450 nm), green (520 nm) and red (660 nm) wavelengths. 183
The mean flight altitude was 260 m abo ve ground level, in a mountainous area with elevation 184
ranging from 1 ,200 to 1 ,600 m. Both front and side overlap s between fl ying tracks were set to 185
a minimum of 80% each , resulting in a dataset of 1 ,393 images . The images were processed 186
with Pix4D mapper P ro software that combines a structure from m otion (SfM) technique and 187
photogrammetric principles. The result ing orthophoto mosaic, used in this study , covers an 188
area of 1 ,174.61 ha and the average ground sampling distance was 0.15 m (Figure 2). The 189
mean re ported geolocation accuracy was 0.022 m in horizontal and 0.060m in vertical 190
direction. 191

10

192
Fig. 2. Orthophoto mosaic based on images captured by Canon S110 RGB camera with eBee RTK 193
drone . 194
195
2.5 Methods 196
When comparing the ALS, UAV and field reference datasets, it is assumed that the last one 197
would be the most accurate. Therefore, the corner of each plot recorded by the Geo XH 198
Trimble receiver was differentially post -processed using the closest EUREF (Euro pean 199
reference frame) station that provides both GPS and GLO NASS (Globalnaya navigatsionnaya 200
sputnikovaya Sistema) RINEX (Receiver Independent Exchange Format) files. We used the 201
files provided by Oroshaza EUREF station, located in Hungary, less than 200 k m from each 202
of the plots. The final positioning accuracy of the measured corner of each plot ranged 203
between 0.8 (plot 38) and 1.6 m (plot 40). 204
2.5.1 ALS data processing 205
The unclassified ALS data were clipped using the limits of each plot, thus allowing the ALS 206
data to be linked to the field measurements. An automatic gr ound classification algorithm 207
based on the Adaptive TIN Ground Filter method of the LP360 software (QCoherent 208
Software LLC, Madison, AL, USA; Weir, 2012) was applied and the ALS points were 209

11
partia lly classified into ground class, while the remaining ALS po ints were reclassified as 210
vegetat ion. The digital terrain model (DTM) was generated from the ground class of the ALS 211
point cloud. Similarly , the digital surface model (DSM) was extracted using the first returns 212
of the vegetation point class. The canopy hei ght model (CHM) for each plot was obtained by 213
subtracting the DTM from DSM. The spatial resolution of the CHM was 0.5 meters. 214
2.5.2 OBIA classification and accuracy assessment 215
To extract the canopy comp osed of spruce and beech trees , a supervised classification was 216
carried out using Object Based Image Analysis (OBIA) embedded within the commercial 217
software eCognition ver. 7. The OBIA technique is based on a segmentation procedure in 218
which, prior to the classification process, the image pixels are grouped into mea ningful 219
objects that correspond to real -world entities by means of a scale factor which limits the size 220
and heterogeneit y of the result ing image objects in conjunction with a homogeneity criterion 221
confined by spectral and shape characteristics (Baatz and S chäpe, 2000). By grouping the 222
pixels into larger meaningful objects, the OBIA approach is particularly suitable in the c ase 223
of very high spatial resolution imagery in which pixels are significantly smaller than the 224
objects of interest (Blaschke et al. 2011 ). 225
The three imagery spectral bands (red, green, blue) were use d together with the rasterized 226
Canopy Height Model for t he segmentation process using the multiresolution segmentation 227
algorithm. The importance (weight) of the CHM band was set three points higher than t he 228
spectral bands. The most suitable segmentation parameters were empirically assessed through 229
visual inspec tion of multiple segmentation attempts. 230
A 2-step segmentation and classification approach was followed: in the first phase , the goal 231
was only to separate the canopy gaps (i.e. bare ground) from the canopy cover (i.e. vegetation 232
area). In this case , the scale parameter was set to 15 and the shape factor to 0.4 (with 0.5 233
compactness). The canopy -gaps were classified by imposing a threshold condition for the 234

12
mean Brightness metric (i.e. (
+
+
)/3). In the second step , the canopy cover class was re – 235
segmented using a scale factor of 25, while for homogeneity criteria , the shape factor was set 236
to 0.6 , indicati ng that the spectral information counted for 0.4 in the decision of pixel 237
aggregation. The most convenient scale parameter was empirically assessed through multiple 238
successive attempts. The initial intention was to segment individual trees crowns but, due to 239
crown overlapping and increased heterogeneity in t he same crowns (as a result of 240
shadowing) , this purpose was not achieved ; hence, we segment the pixels in terms of 241
enclosed parts of both spruce and beech crowns . For plots 40 and 41 , the canopy cover wa s 242
classified in two classes (i.e. beech cr owns and spruce crowns) while i n the case of plot s 38 243
and 44 , an additional class was added belonging to ‘other hardwood ’ species because those 244
plots also contained individuals of birch ( Betula pendula ), ash ( Fraxi nus excelsior ) or 245
sycamore ( Acer pseudoplatanus ). 246
The standard Nearest Neighbour algorithm was adopted for image classification. More than 247
25 object features were defined for the discrimination of the classification categories (i.e. 248
spruce crowns, beech crowns, other hardwoods). We defined various spectral, texture and 249
vertic al structure metrics derived from the UAV imagery, similar to Michez et al. (2016 ). As 250
spectral metrics , we used the mean bands and standard deviation values in the three visible 251
band s, the normalized bands for red (
/(
+
+
)), green (
/(
+
+
)) and blue 252
((
/(
+
+
)), band ratios (
/
) and several spectral indexes: Normalized Green Blue Index 253
– NGBI ((
–
/
+
)), Normalized Red Blue Index – NRBI ((
–
)/(
+
)) and Normalized 254
Gree n-Red Vegetation Index ((
–
)/(
+
)). As shape metrics we tested compactness, 255
density, roundness and skewness of the height (CHM). For the textural measure , we used 256
Grey Level Co -occurrence Matrix (GLCM) (h omogeneity, standard deviation), which is the 257
most common ly used metric for te xtural evaluation (Spracklen and Spracklen , 2019) . 258

13
The Feature Space Optimization tool embedded within eCognition was employed in order to 259
semi -automatic ally choose the optimal obj ect features (metrics) that allowed the best 260
delineation between the targeted classes. 261
The accuracy of the classification was performed by generating a standard confusion matrix. 262
All the trees crowns wi thin the four experimental plots were manually digiti zed through 263
visual interpretation of the RGB imagery. The field reference data (i.e. trees positions, 264
species and crown projection) were overlaid on the imagery and used to assist the photo – 265
interpretatio n process. The manually digitized crowns were used as a reference for the 266
accuracy assessment of OBIA classification. 267
2.5.3 Individual tree identification 268
To detect individual trees based on ALS data, the derived CHM was divided in two separate 269
CHMs, one for each tree species (i.e. Norway spruce and beech). T his delineation operation 270
of the CHM was performed for each plot using the canopy areas covered with Norway spruce 271
and beech trees (i.e. as r esulted from OBIA classification). 272
The individual trees were detected based on the two CHMs using the Canopy Maxim a 273
algorithm implemented in the Fusion software (McGaughey, 2018). This algorithm uses the 274
correlation between crown width and tree heights (P opescu et al. 2002; Popescu and Wynne, 275
2004) to identify the local maxima and the crown width of a tree. 276
In the ca se of Romania, there have been no standardized equations that could be used for 277
expressing the variation of crown diameter according to the h eight of the trees or for each tree 278
species considered (Norway spruce and beech). Thus, it was necessary to build s uch local 279
equations based on crown diameter and tree height field measurements for the two species. 280
During the field inventory , the crown pro jection area (Pr crown) was measured, with the crown 281
diameter (d crown) being computed based on a circular shape of t he crown projection area. 282

14
The Canopy Maxima algorithm created a file with the coordinates, height and crown width of 283
each individual identif ied tree. The detected trees within each plot were visually linked with 284
the field measured trees according to the p ositions of nearest neighbouring trees. 285
The omission and commission rate errors were provided at plot level. The omission error is 286
given by the number of detected trees that cannot be matched with a reference tree. The 287
commission error is given by the tot al number of trees detected by the canopy maxima 288
algorithm and not matched with field inventory trees. 289
The ground reference tree volume (v i) of each tree species (spruce and beech) was calculated 290
using a double logarithmic equation with tree diameter at th e breast height (dbh) and height 291
(h) as variables (eq. 1) along with specific regression coefficients for our country (Giurgiu et 292
al. 2004 ). 293
(1) 294
The total volume (per hectare) was calculated by summing the volumes of all individual tree 295
species. 296
2.5.4 Tree dbh and volume estimations, best model selection and cross -validation 297
Because the dbh is not directly extracted based on ALS data, a linear regression model was 298
used to develop local equations relating ALS derived height and crown diameter fo r each tree 299
species with the field measured dbh. 300
To choose the right model to est imate dbh value based on the ALS measurements of tree 301
height and crown width, a cross validation algorithm using model functions from the 302
tidyverse package (Wickham, 2017) in R software (Team, 2018) was employed . The 303
simulations are based on a Monte Carlo algorithm (Chattopadhyay and Chattopadhyay, 304
2014) . The algorithm relies heavily on repeated random sampling. A number of 25 random 305
splits were performed with percentage s of 80 and 20 for training and test data , respec tively . A 306
larger number of simulations was conducted but the results were close to the ones obtain ed 307

15
from only 25 simulations. Five different models from two regression families ( linear and 308
nonlinear regression s) were chosen to find the best fit. For every given model, an estimation 309
of its performance was made through the aggregation of its performance in each of the splits. 310
For the chosen model, we made an estimate regarding how much confidence we should have 311
in the significance of each of its coefficients . 312
In three of the plots (41, 44 and 40) , Norway spruce is the dominant species whereas one plot 313
(38) is dominated by beech. The three plots dominated by Norway spruce were considered as 314
having the same dbh dynamic , and a set of simulation s was conducted fo r each of the two 315
species. Plot 38, where beech is the main species , was considered to have a different dbh 316
dynamic for each of the species compared with the other three plots. Separate simulations 317
were conducted for this plot on each of the species. 318
To compare the performance of models on their training sets with their performance on their 319
test sets , a qualitative assessment of the kernel density of the residuals was conducted. Also, 320
root-mean -squared error was computed because it provides a useful first idea of how a given 321
model performs, and it is expressed in the same units as the response variable. 322
Using the predicted dbh values and height obtained with ALS, tree volume was computed 323
based on equation (1) along with specific regression coefficients for o ur country (Giurgiu et 324
al. 2004 ), and an estimation of their accuracy was made. 325
To test the significance difference betwe en the mean dbh for the population of the reference 326
field trees with the respect to the mean dbh for the population of the matched ide ntified ALS 327
trees as well as evaluating whether the two sets of variables follow normal or nonparametric 328
distribution, the paired sample t -test or Wilcoxon test was used for each plot and for each tree 329
species separately. We performed the same method for the case of the mean volume for the 330
population of the reference field trees with the respect to the mean volume for the population 331
of the matched identified ALS trees. 332

16
To measure the deviations of the ALS -estimated variables (tree height, crown diameter, d bh 333
and tree volume) from the matched field variables, the root mea n square error (RMSE) (eq. 2) 334
was determined for each plot, variable and tree species. An overall RMSE was reported for 335
each tree species. For the pre dicted variables (dbh and tree volume) , the relative root mean 336
square error (rRMSE) (eq. 3) was determined. 337
(2), 338
(3), 339
where
is the tree species ’ variables (tree height, cr own diameter, dbh and tree volume) and 340
is the average value of the tree dbh and volume measured variables. 341
The workflow that provides the general methodology applied in the current study is shown in 342
Figure 3. 343

17

344
Fig. 3. Overview of the workfl ow methodology . 345
346
3. Results 347
3.1 OBIA classification and accuracy assessment 348

18
The overall accuracy for OBIA classification (Table 3) was between 73.9% (plot 38) and 349
77.3% (plot 41). The results of the OBIA classification are presented graphically as polygons 350
overla id with the corresponding RGB imagery in Figure 4. 351
The user’s accuracy (UA) was generally higher than producer’s accuracy (PA) for all the 352
plots and all the classification classes, except for the canopy gaps class for which the PA was 353
higher. The cla ssific ation of spruce crowns yielded a UA of 71.1% to 81.8% and a PA of 354
68.9% to 77.9% being close to that achieved for beech crowns , which had a UA of 62.2% to 355
87.8% and a PA between 60.5% and 78.9%. 356
In the case of spruce, both omission and commission er rors were mostly due to the confusion 357
with canopy gaps class while in case of beech , the errors were relative equally distrib uted 358
among the spruce and canopy gaps class. 359
Table 3. Accuracy assessment of OBIA tree species’ classification. 360
Plot
Spruce Beech Canopy -gaps Other species
PA/UA %
PA/UA % PA/UA % PA/UA % Ovr. %
38 77.9/71.0 71.9/87.8 77.2/55.2 52.4/67.2 73.9
40 74.7/81.2 60.5/62.2 82.2/74.8 NA 75.9
41 68.9/81.5 78.9/84.9 86.1/66.7 NA 77.3
44 69.7/81.8 67.9/77.1 89.0/71.7 78.1/84.8 76.3
Note: UA, User’s accuracy; PA, Producer’s accuracy; Ovr, Overall accuracy 361

a)

b)

19

c)

d)
Fig. 4. The polygons resulted from OBIA classification overlaid on the UAV RGB imagery (Magenta 362
– spruce, Green – beech, Other species – yellow – a) plot 41, b) plot 4 4, c) plot 38, d) plot 40. 363
364
3.2 Individual tree detection based on ALS CHM 365
The Canopy Maxima specif ic parameters computed locally based on field measurements for 366
each tree species are presented in Table 4. 367
Table 4. Canopy maxima specific parameters (A, B, C, D) computed locally for tree species within each plot. 368
Plot Species Canopy maxima specific parameters
A B C D
41 Norway spruce 3.987 -0.0694 0.004 0
Beech 2.5072 0.0186 0.0031 0
44 Norway spruce 4.2891 -0.0681 0.0037 0
Beech 6.0432 -0.3933 0.0125 0
38 Norway spruce 3.4336 -0.0585 0.0041 0
Beech 3.6464 -0.2004 0.0086 0
40 Norway spruce 2.5422 -0.0176 0.0038 0
Beech 3.248 -0.0962 0.0073 0
369
Taking into account the upper canopy (trees included in Kraft class I, II and III), the detection 370
rate in the case of Norway spruce trees was between 57% (plot 40) and 73% (plot 44). In the 371

20
case of beech tree species, the rate detection was between 19% (pl ot 38) and 37% (plot 40). 372
The omission/commission rate errors are provided at plot level (Figure 5). 373

a)

b)
Fig. 5. Omission (a) and commission errors (b) for each plot. 374
The low detection tree rate of the beech trees could be attributed to the fact that the canopy 375
maxima algorithm used does not perform well for deciduous trees, which have more rounded 376
crown shapes compared with coniferous trees and the crowns tend to overlap near the top of 377
the tree (McGaughey, 2018), as in our case. Results of ALS individual tree detection, 378
matched tree tops and estimated tree crowns for both tree species are shown in F igure 6. 379

a)

b)

21

c)

d)
Fig. 6. ALS matched tree tops, estimated tree crowns for Norway spruce and beech within each plot 380
overlaid on UAV RGB imagery a) plot 41, b) plot 44, c) plot 38, d) plot 40. 381
382
For the matched trees, we calculated separately f or each tree species and each plot the mean 383
height of the field -measured trees (
) and the mean height of the automatic ally identified 384
ALS matched trees (
), respectively the mean crown diameter of the field -measured trees 385
(
) and the mean height of the automatically identified ALS matched trees 386
(
) (Table 5). RMSE values were calculated for both variables (tree height and 387
crown diameter). 388
Table 5. Mean height, crown diameter values of field -measured trees an d matched ALS -derived mean height and crown 389
diameter values. 390
Plot Tree species Matched
trees

(m)

(m)

(m)

(m)

(m)

(m)
41 Norway spruce 96 34.37 34.92 1.41 6.44 3.15 3.69
Beech 28 31.71 32.19 1.23 5.73 3.10 2.93
44 Norway spruce 118 32.99 33.37 0.77 6.06 3.64 2.89
Beech 17 29.58 29.74 1.02 4.59 3.12 1.69

22
38 Norway spruce 101 31.34 30.95 0.71 5.70 3.73 2.39
Beech 67 27.91 27.96 0.86 4.85 2.99 2.17
40 Norway spruce 156 30.15 29 1.27 5.63 3.33 2.55
Beech 34 24.2 23.46 1.63 5.47 2.07 3.59
All
plots Norway spruce 471 31.98 31.73 1.09 5.95 3.47 2.87
Beech 146 27.98 27.93 1.17 5.13 2.81 2.67
391
In the case of spruce trees and for each plot, strong and significant correlations were obtained 392
(α <5%) (r=.917 -.972) between field-measur ed tree heights and matched tree heights of 393
identified ALS trees (Figure 7). A significant correlation was also found for each plot (α 394
<5%) (r=.250 -.426) between the field crown calculated diameter and the matched crown 395
diameter of identified A LS trees. 396

a)

b)

c)

d)
Fig. 7. Lower panel represents scatter plots for all field reference measured/computed variables of 397
Norway spruce trees ( Crown.diameter_field , Height_field , dbh_field and Volume_field ) and ALS – 398
estimated ( Height_ALS , Crown.diameter_ALS) ones. The upper panel represents the correlation 399
coefficients(r) and their significance (**p<0.01; *p<0.05). Distribution of the experimental values is 400
presented in the diagonal line a) plot 41, b) plot 44, c) plot 38, d) pl ot 40. 401

23
In the case of beech tre es and for each plot, the height correlations were also strong and 402
significant (α <5%) (r=.767 -.935) (Figure 8). Similarly, significant correlations were found (α 403
<5%) between the field crown calculated diameter and the match ed crown diameter of 404
identified ALS trees (r=.250 -.426). 405

a)

b)

c)

d)
Fig. 8. Lower panel represents scatter plots for all field reference measured/computed variables of 406
beech trees ( Crown.diameter_field , Height_field , dbh_field and Volume_field ) and ALS -estimated 407
(Height_ALS , Crown.diameter_ALS) ones. The upper panel represents the correlation coefficients(r) 408
and their significance (**p<0.01; *p<0.05). Distribution of the experimental values is presented in the 409
diagonal line a) plot 41, b) plot 44, c) plot 38, d) plot 40. 410
For both tree species , the crown diameter derived from ALS data would be , in our case, 411
underestimated compared with the crown diameter calculated from the field measurements. 412
The algorithm that calculates the crown diameter for the ALS CHM does not take into 413
account the overlapping crown diameters (Popescu, 2007) of the two species, but would 414
instead have the overlapping crown diameters considered when the tree crown projections are 415
measured during the field inventory. 416
417

24
3.3 Tree dbh and volume estimations, best mo del selection and cross -validation 418
Regarding tree dbh, we can consider this along with tree height as the two most important 419
elements for t he estimation of tree volume. Despite tree dbh not being directly mea sured on 420
the CHMs derived from ALS data, the A LS extracted height and crown diameter were found 421
to be strongly correlated with the dbh measured in the field, for both Norway spruce and 422
beech species (Figure s 7 and 8 ). 423
As a result of Monte Carlo simulations, t he best prediction model for the dbh was ch osen 424
according to the lowest RMSE obtained for the testing set (Figure 9, Table 5). 425

a)
b)
Fig. 9. Best model for dbh prediction — results of Monte Carlo simulation a) Norway spruce b) 426
beech. 427

25
The dbh regression analysis for all spruce trees demonstr ated an RMSE range between 4.8 428
and 6.1cm (Table 5). The overall RMSE calculated for the spruce trees within all plots was 429
5.8 cm, which represents 12% of the mean dbh value of field linked measured spruce t rees, 430
and 14% of the mean dbh of all spruce trees within all plots. 431
Compared with the spruce trees, the results of the dbh regression analysis for the beech trees 432
showed an RMSE range between 5.6 and 6.2 cm, slightly higher than in the case of spruce 433
(Tab le 5). The overall RMSE calculated for the beech t rees within all plots was 5.9 cm (20% 434
of the mean dbh value of field linked measured beech trees, and 25% of the mean dbh of all 435
beech trees within all plots) (Table 6). 436
437
Table 6. Linear regression for the prediction of dbh based on ALS data . 438
Plot Tree spe cies Dependent variable
(field measurement) Model and significant variables :
ALS derived height (H), ALS
derived crown dimeter (CrD) R2 RMSE _
dbh
(cm)
38 Norway spruce dbh (cm) 2.003*H+1.523* CrD-21.053 0.62 4.8
Beech 0.960*H+1.490*CrD -1.992 0.17 6.2
41,44,40 Norway spruce dbh (cm) 1.675 *H+1.5 04* CrD -9.345 0.56 6.1
Beech 0.624*H +2.183* CrD+6.597 0.42 5.6
439
Comparing the mean dbh of the field trees for each tree species and the mean predicted dbh 440
based on ALS data, no statistically significant diffe rences were found (p>0.05) (Table 7). 441
Table 7. Mean dbh measured in the field and mean predicted dbh based on ALS data for the same tr ees and T -paired, 442
Wilcoxon test statistical significance s. 443
Plot Tree
species Number
of
matched
trees
Fielddbh(cm)
ALSdbh(cm) Freedom
degree Tvalue,
0.05 pvalue Wilcoxon
Zvalue Wilcoxon
pvalue
38 Norway
spruce 201 46.616 46.619 200 -0.002 0.998
beech 66 28.912 28.912 -0.687 0.492
41,44,
40 Norway
spruce 372 49.25 49.25 -0.288 0.774
beech 79 29.82 28.83 -1.676 0.094
444

26
The measured dbh of the field trees and the ALS -predicted dbh for each tree species are 445
presented in Figure 10. 446

a)
b)
Fig. 10 . Field -measured dbh and predicted dbh based on ALS data for a) Norway spruce; b) beech . 447
448

27
The previously predicted dbh and ALS -derived height were used together as inputs for the 449
volume estimation of each tree species. To estimate tree volume, equation (1) was used along 450
with specific regression coefficients for our country (Giurgiu et al. 2 004). 451
In relation to the mean volume of trees estimated based on ALS data compared with the mean 452
volume of the field -linked trees for each tree species (spruce and beech), no statistically 453
significant differences were detected (p>0.05) (Table 8). 454
Table 8. Mean volume of trees calculated based on field measurements and mean estimated volume based on ALS data — 455
Wilcoxon test and statistical significance. 456
Plot
Tree
species Number of
matched
trees
Fieldv(m3)
ALSv (
m3) Wilcoxo n
Zvalue Wilcoxon
pvalue
38 Norway
spruce 100 2.29 2.24 -.087 0.930
beech 66 0.97 0.92 0.457 0.648
41,44,40 Norway
spruce 372 2.59 2.56 -0.550 0.582
beech 79 1.07 0.97 1.933 0.053
457
The ALS estimated tree volume for each of the tree species and the calculated tree volume of 458
the field reference trees are presented in Figure 1 1. 459
a)

28

b)
Fig. 11. Field-calculated volume of trees and estimated volume of trees based on ALS data for a) 460
Norway spruce; b) beech . 461
462
The RMSE of the spruce trees’ volume estima tion based on ALS data ranged between 0.4 m3 463
and 0. 6 m3 depending on the plot, while the RMSE of the beech trees’ volume estimation 464
based on ALS data yielded slightly lower values, between 0.3 m3 and 0. 5 m3 depending on 465
the plot. The overall RMSE f or the spruce trees’ volume was 0.5 m3 , which represents 20% 466
of the mean volume of field -linked measured spruce trees, and 29% of the mean volume of 467
all spruce trees within all plots, while the RMSE of the beech trees’ volume estimation based 468
on ALS data was 0. 4 m3, representing 42% of the mean volume of field -linked measured 469
beech trees, and 55% of the mean volume of all beech trees within all plots. 470
The total wood volume of the identified trees by ALS data represents up to 79% (Norway 471
spruce) and 29% (be ech) o f the total stand volume of the two species, calculated based on 472
field measurements (Table 9). 473
Table 9. Percentage error of estimated total volume based on ALS data from total field reference volume . 474
Plot Tree
species Total
ground
reference
volume
(m3) ALS
estimated
volume of
matched
trees
(m3)
Field reference
volume of ALS –
matched trees
(m3) Percentage of the
total ALS -estimated
volume
compared to
total field
reference
volume RMSE of
volume
(m3)

29
(%)
0 1 2 3 4 5

6
41 Norway
spruce 438.1 317.7 304.4 73 0.6
Beech 180.0 38.6 39.5 21 0.5
44 Norway
spruce 433.0 342.7 342.6 79 0.6
Beech 79.6 19.0 22.6 24 0.5
38 Norway
spruce 298.7 224.0 229.2 75 0.4
Beech 210.3 61.4 64.6 29 0.4
40 Norway
spruce 504.0 292.8 316.6 58 0.4
Beech 88.1 17.2 23.0 20 0.2
475
4. Discussion 476
There have been numerous studies combin ing ALS with different types of data, which were 477
provided by satellite or aerial platforms that possess optical, hyperspectral or SAR sensors. 478
This study explored the use of field measurements combi ned with d erived data from ALS and 479
UAV aerial imagery (high -resolution orthoimagery processed by computer vision software) 480
because this approach is only in the early stages of development but is perceived as having 481
significant potential for forestry applic ations such as obtaining canopy attributes (Chianucci 482
et al. 2015). UAV technology (platforms, sensors, data and software) is still under a period of 483
highly dynamic development and promises to provide foresters and researchers with a field – 484
portable remote sensing device for real -time applications that offers a low -cost option for the 485
collection of high -precision 3D data when and where it is required . 486
While there are also a large number of studies that have been dedicated to forest assessment 487
using ALS, mos t of these works have focused on bo real forests dominated by coniferous trees 488
that typically show more distinct gaps between single tree crowns. The present study extends 489
the use of ALS technology in mixed forests, being located in a mixed temperate forest stand 490
characterized by a dense canopy structure and an irreg ular mixture of species (spruce, beech 491
and other deciduous trees). 492

30
Our OBIA classification of crown species may be considered sufficiently satisfactory 493
considering that we used single RGB imagery and not more expensive multispectral or 494
hyperspectral datase ts. We took advantage however of a n ALS -generated Canopy Height 495
Model obtained from a previous flight which contributed in improving the segmentation 496
results. Also, we applied a simple and fast classification method (i.e. Standard Nearest 497
Neighbour) while currently gaining in popularity are the more sophisticated machine learning 498
approaches such as random forest (Michez et al. 2016, Nevalainen et al. 2017, Franklin et al. 499
2017) due to their good p erformances and capacity to handle large amount s of heterogen ous 500
data. 501
Michez et al. (2016) used the OBIA method in conjunction with a random forest algorithm to 502
classify riparian forest tree species for two Wallonia n sites and obtained an overall 503
classif ication accuracy of 79.5% and 84.1% for five targeted forest classes. They used 504
however a multi -temporal dataset of RGB and RGNIR UAV imagery (i.e. eighteen for one 505
site and seven for the other) , an approach which significantly improved species classificat ion. 506
An overall classification accuracy of 78% was achieved by Franklin et al. (2017) on 6-band 507
UAV imagery acquired over a hardwood forest in Ontario. The multiresolution segmentation 508
process was used to generate image objects that corresponded to indivi dual tree crowns while 509
a random forest algorith m was used to separate the species based on spectral, textural and 510
crown shape variables. Their goal was however to distinguish between four deciduous 511
species (aspen, birch and two maple species) , a strategy which requires a finer differentia tion 512
than classifying a deciduous species from a resinous one. 513
Vastaranta et al. (2011) ran Monte Carlo simulations to investigate how the errors related to 514
detection of trees, tree height and tree diameter prediction influ ence individual trees ’ 515
inventories. They co ncluded that the accuracy of forest inventory parameters ’ retrieval at tree 516
level based on ALS data are particularly dependent on the tree detection rate which was 517

31
found be the most important source of errors. Tre e height is the most acc urate variable 518
retrieved from ALS data while dbh must be modelled from height and/or crown diameter , but 519
the relationship between these three parameters is highly variable and sensitive due to errors 520
of input data. In our study, the detection rate of individual tree s based o n the Cano py Maxima 521
algorithm and ALS CHM was up to 73% (plot 4 4) for the Norway spruce trees and up to 37% 522
(plot 40) in the case of beech trees. Consequently, we obtained a high rRMSE for beech 523
predicted variable s (e.g. 59.9% RMSE for volume, 24.54% RMSE for dbh). 524
Thiel and Schmullius (2016) used a CanopyMaxima Fusion implemented algorithm, similar 525
to our study, to compare the tree detection rate on canopy height models obtained from 526
LiDAR and UAV data on a 4 -hectare mixed forest (mainly spruce and pine) in central -east 527
Germany. They reported a detection rate of 78.0% for LiDAR CHM and 93.2% for UAV – 528
derived CHM, while the commission errors were approximately 10%. The higher detection 529
rate may be attributed in part to a finer resolution of the ALS -derived CHM (i.e. 25 X 25 cm) 530
as well as for the RGB imagery (i.e. 10 cm). 531
In our study, we used a single -tree approach to estimate the individual tree volume and total 532
volume of the spruce and beech trees within each plo t. The correlation between canopy cover 533
density/number of correctly detected trees has been highlight ed in previous studies. 534
Falkowski et al. (2008) demonstrated that in dense forest, it is particularly difficult to find 535
algorithms to discriminate at a sat isfactory rate between crowns while Wang et al. (2016) 536
found that canopy vertical structure plays a c onsiderable role in the detection accuracy, and 537
has increased impact even compared to tree species or species composition in a stand. Our 538
model, which made use of the CHM, provides best results for even -aged mixed stands where 539
the trees’ vertical structure is less varied. 540
We achieved an overall RMSE for tree height of 1.09 m for Norway spruce and 1.17 m for 541
beech, results which are consistent with other res earch reports. Recent studies for Norway 542

32
spruce pure stands in Romania (Apostol et al. 2016) obtained an RMSE of 1.7 –2.2 m for 543
height of Norway spruce trees using the same individual tree detection method as in the 544
present study. In a biomass study in Cost a Rica, using UAV -based point clouds, in an area 545
covered with tropical forests but without species di scrimination, an RMSE of 0.87 m for tree 546
height was noted (Zahawi et al. 2015). In a sparse eucalypt forest in south -eastern Australia, 547
using UAV -SfM and A LS data, for tree height results an RMSE of 0.92 –1.3 m was reported 548
(Wallace et al., 2016). Older stu dies with coniferous and deciduous forests (Popescu and 549
Zhao, 2008) also found an RMSE for a height of 1.38 –2.15 m. 550
In a study regarding Sacramento street trees using ALS data and aerial ortophotoimagery, the 551
RMSEs obtained were 1.64 m for tree height, 1.07 m for crown diameter and 10.32 cm for 552
DBH derived from crown diameter (Lee et al. 2016). 553
Crown diameter and height accounted for a maximum of 62% (plot 3 8) of dbh variance of 554
field-measured spruce trees , with an RMSE of 5.8 cm . Popescu (2007) also found for pine 555
trees that lidar -estimated height and crown diameter explained 87% of dbh variance, with an 556
RMSE of 4.9 cm , better with 0.9 cm than the one obtain ed in our study for spruce trees . Other 557
studies involving Norway spruce stands and Scots p ine from Sweden (Persson et al. 2002) 558
demonstrated an RMSE for dbh estimation of 3.8 cm, better with 2 cm than the one obtained 559
in our study. 560
The RMSE between field -calculated and ALS -derived crown diameter was 2.87 (Norway 561
spruce) and 2.67 m (beech). Falk owsky et al. (2008) obtained an RMSE of 1.29 –2.02 m for 562
the crown diameter in a conifer forest, depending on canopy cover conditions and method 563
utilized (i.e. spatial wavelet analysis and variable window filters). Similar results were 564
obtained by Forzieri et al. (2009) with RMSE for crown diameter attaining values from 1.43 565
to 6.85 m, while Popescu and Zhao (2008) achieved a range of 1.84 –2.08 m. Kato et al. 566
(2009) als o found that the crown width between LiDAR and field measurements were highly 567

33
correlated, with an RMSE of 0.93 m and 2.89 m for coniferous and deciduous species, 568
respectively. In an individual tree detection study based on ALS data in tropical forests of 569
Panama, the models offered an RMSE of 1.23 m (Ferraz et al. 2016). 570
The RMSE of dbh field -measured and ALS -estimated data was 5.8 cm (Norway spruce) and 571
5.9 cm (beech), close to results obtained by Vauhkonena et al. (2010) with an RMSE for dbh 572
of 6.3 cm (Norway spruce ). 573
In terms of the LiDAR estimation volume, where a comparison is conducted betw een the 574
volume of the trees measured at the ground with the corresponding ALS -estimated volume of 575
the same trees, RMSE values of 0.5 m3 (Norway spruce) and 0.4 m3 (beech) were recorded. 576
Apostol et al. (2016) obtained for Norway spruce pure stands, using on ly the ALS -derived 577
height as input for volume estimation, an RMSE ranging from 0.5 to 0.7 m3. Vauhkonena et 578
al. (2010) produced an RMSE for the volume of trees of 0.55 m3 (Norway spruce). On the 579
other hand, Yao et al. (2012) obtained RMSE values for the sa me parameter ranging from 580
0.43 to 0.46 m3 for coniferous and 0.50 to 0.60 m3 for deciduous trees. 581
5. Conclusions 582
In this study , we prop osed a workflow to detect individu al trees based on ALS and UAV data 583
in mixed Norway spruce -beech even -aged plot stands. An OBIA classification was conducted 584
in order to delineate the spruce and beech crown cover trees and canopy gaps within all plots. 585
The overall accuracy assessment of th e OBIA classification demonstrated that by using RGB 586
imagery , the crown cover of the two t ree species could be delineated with high accuracy. 587
The individual tree detection algorithm used together with ALS CHM revealed a high 588
detection rate for the spruce t rees, compared to the detection rate for the beech trees. The 589
relatively low detection tre e rate of the beech trees could be attributed to the fact that the 590
algorithm used does not perform well for the mixed stands of deciduous trees (i.e. beech) and 591
spruc e tree species, as was the case for our study. 592

34
The regression model for the dbh prediction used as the significant variables the ALS -derived 593
crown diameter and height of the tree. The overall RMSE for the dbh yielded similar values 594
for both tree species. M oreover, no statistically significant difference (p > 0.05) was revealed 595
between the ALS -predicted and field-measured dbh. 596
The individual tree detection method reported in this study provide a high accuracy for the 597
estimation of the total spruce trees volume, while the total volume of the beech trees is less 598
accurate estimated. Further investigations are required, particularly with respect to choosing 599
the best method for individual tree detection of beech trees located in temperate even -aged 600
mixed stands. 601
602
Acknowledgments 603
This work was supported by the Romanian Ministry of Resea rch and Innovation within the 604
National “Nucleu ” Program — projects codes PN09460118, PN09460113, PN16330107, 605
PN18040102, PN19070109. 606
607
References 608
609
Apostol, B., Lorenț, A., Petrila, M. , Gancz, V., Badea, O. (2016). Height extraction and stand 610
volume estimati on based on fusion airborne LiDAR data and terrestrial measurements 611
for a Norway spruce [Picea abies (L.) Karst.] test site in Romania. Notulae Bo tanicae 612
Horti Agrobotanici Cluj -Napoca, 44(1), 313 -326. http://doi.org/10.15835/nbha44110155 613
Baatz M, Schäpe M. (2000). Multiresolution segmentation — an optimization approach for 614
high quality multi -scale image segmen – tation. In: Strobl J, Blaschke T, Griesebner G 615
(Eds), Angewandte Geographische Informations -Verarbeitung XII. Wichmann Verlag, 616
Karlsruhe, 12 –23. 617
Blaschke T, Johansen K, Tiede D. (2011). Object -Based Image Analysis for Vegetation 618
Mapping and Monitoring. In: Weng Q., Advances in Environmental Remote Sensing: 619
Sensors, Algorithms, and Applications, 141 -266 620

35
Chattopadhyay, A.K., Chattopadhyay, T., (2014 ). Monte Carlo simulation, in: Springer Series 621
in Astrostatistics. https://doi.org/10.1007/978 -1-4939 -1507 -1_10 622
Chianucci, F., Disperati, L., Guzzi, D., Bianchini, D., Nardino, V., Lastri, C., Rindinella, A. , 623
Corona, P. (2016). Estimation of canopy attribut es in beech forests using true colour 624
digital images from a small fixed -wing UAV, International Journal of Applied Earth 625
Observation and Geoinformation, 47, 60 –68. 626
Dalponte, M., Ørka, H. O., Ene, L. T., Gobakken, T., Næsset, E. (2014). Tree crown 627
delineati on and tree species classification in boreal forests using hyperspectral and ALS 628
data. Remote S ensing of Environment , 140, 306 –317. 629
https://doi.org/10.1016/j.rse.2013.09.006 630
Duncanson, L.I., Cook, B.D., Hurtt, G.C. , Dubayah, RO. (2014). An efficient, multi -layered 631
crown delineation algorithm for mapping individual tree structure across multiple 632
ecosystems. Remote Sensing of Environment 154, 378 -386. 633
Edson, C., Wing, M.G. (2011). Airborne light detection and ranging (LiDAR) for individual 634
tree stem location, height, and biomass measurements . Remote Sensing 3, 2494 -2528, 635
doi:10.3390/rs3112494. 636
Falkowski, M.J., Smith, A.M.S., Gessler, P.E. , Hudak, A.T., Vierling, L.A. , Evans, J.S. 637
(2008). The influence of conifer forest canopy cover on the accuracy of two individual 638
tree measurement algorithms usi ng lidar data. Canadian Journal of Remote Sensing, 639
34(Suppl. 2), S338 -S350. 640
Ferraz, A., Bretar, F., Jacquemoud, S., Gonçalv es, G., Pereira, L., Tomé, M., Soares, P. 641
(2012). 3 -D mapping of a multi -layered Mediterranean forest using ALS data. Remote 642
Sensing of Environment, 121, 210 –223. http://d oi.org/10.1016/j.rse.2012.01.020 643
Ferraz, A., Saatchi, S., Mallet, C. and Meyer, V. (2016) . Lidar detection of individual tree 644
size in tropical forests, Remote Se nsing of Environment . Elsevier Inc., 183, 318 –333. 645
doi: 10.1016/j.rse.2016.05.028. 646
Forzieri, G. , Guarnieri, L., Vivoni, E.R., Caste lli, F., Preti, F. (2009). Multiple attribute 647
decision making for individual tree detection using high -resolution laser scann ing.Forest 648
Ecology and Management 258 (11), 2501 -2510. DOI: 10.1016/j.foreco.2009.09.006 649
Frank lin, S. E., Ahmed, O. S. (2017). Deciduous tree species classification using object -based 650
analysis and machine learning with unmanned aerial vehicle multispectra l data. 651
International Journal of Remote Sensing, 00(00), 1 –10. 652
http://doi.org/10.1080/01431161.2017.1363442 653
Gatziolis, D., Lienard, J. F., Vogs, A., Strigul, N. S. (2015). 3D tree dimensionality 654
assessment using photogrammetry and small unmanned aerial vehicles. PLoS ONE, 655
10(9), 1 –21. http://doi.org/10.1371/journal.pone.0137765 656
Giurgiu, V., Decei, I., Drăghiciu, D. (2004). Metode și tabele dendrometrice [Forest 657

36
Mensuration Methods and Tables]. Bucharest, Ed Ceres (in Romanian). 658
Heinzel, J. N., Weinacker, H., Koch, B. (2008). Full automatic detection of tree species based 659
on delineated single tree crown s – a data fusion approach for airborne laser scanning data 660
and aerial photographs. Silvilaser 2008 Proceedings , (2007) , 76–85. 661
Hyyppä, J., Holopainen, M., Olsson, H. (2012). Laser scanning in forests. Remote Sensing , 662
4(10), 2919 –2922. http://doi.org/10.3390/rs4102919 663
Hyyppa, J., Hyyppa, H., Le ckie, D., Gougeon, F., Yu, X., Maltamo, M. (2008). Review of 664
methods of small -footprint airborne laser scanning for extracting forest inventory data in 665
boreal forests. International Journal of Remote Sensing, 29(5), 1339 –1366. 666
http://doi.org/Doi 10.1080/01431160701736489 667
Hyyppä, J., Kelle, O., Lehikoinen, M., Inkinen, M. (2001). A segmentation -based method to 668
retrieve stem volume estimates from 3 -D tree height models produced by laser scanners. 669
IEEE Transactions on Geoscience and Remote Sensing, 39(5), 969 –975. 670
http://doi.org/10.1109/36.921414 671
Kaartinen, H., Hyy ppä, J., Yu, X., Vastaranta, M., Hyyppä, H., Kukko, A., Holopainen, M., 672
Heipke, C., Hirschmugl, M., Morsdorf , F., Næsset, E., Pitkänen, J., Popesc u, S., 673
Solberg, S., Wolf, B.M., Wu, J.C. (2012). An international comparison of individual tree 674
detection and extraction using airborne laser scanning. Remote Sensing, 4(4), 950 –974 675
Kankare, V., Räty, M., Yu, X., Holop ainen, M., Vastaranta, M., Kantola, T., … Viitala, R. 676
(2013). Single tree biomass modelling using airborne laser scanning. ISPRS Journal of 677
Photogra mmetry and Remote Sensing, 85, 66 –73. 678
https: //doi.org/10.1016/j.isprsjprs.2013.08.008 679
Kato, A., Moskal, L. M., Schiess, P., Swanson, M. E., Calhoun, D., Stuetzle, W. (2009). 680
Capturing tree cr own formation through implicit surface reconstruction using airborne 681
lidar data. Remote Sensing of Environme nt, 113(6), 1148 –1162. 682
http://doi.org/10.1016/j.rse.2009.02.010 683
Kraft, G. (1884). Beiträge zur Lehre von Durchforstungen, Schlagstellungen und 684
Lichtungshieben [On the methodology of thinning s, shelterwood cuttings and heavy 685
release operations]. Hanover. 686
Lee, J. H., Ko, Y., McPherson, E. G. (2016). The feasibility of remotely sensed data to 687
estimate urban tree dimensions and biomass. Urban Forestry and Urban Greening, 16, 688
208–220. https://doi.org/10.1016/j.ufug.2016.02.010 689
McGaughey, J. (2018). FUSION/LDV: Software for LiDAR data analysis and visualization. 690
August 2018 – Fusion version 3.80 691
Michez, A., Piégay, H., Lisein, J., Claessens , H., Lejeune, P. (2016). Classification of riparian 692
forest species and health condition using multi -temporal and hyperspatial imagery from 693
unmanned aerial system. Environmental Monitoring and Assessment, 188(3), 1 –19. 694

37
http://doi.org/10.1007/s10661 -015-4996 -2 695
Nevalainen, O., Honkavaara, E., Tuominen, S., Viljanen, N., Hakala, T., Yu, X., Hyyppä, J., 696
Saari, H., Pölönen , I., Imai, N. N., Tommaselli, A. M. G. (2017). Individual tree 697
detection and classification with UAV -Based photogrammetric point clouds and 698
hyperspectral imaging. Remote Sensing, 9(3). http://doi.org/10.339 0/rs9030185 699
Persson, Å, Holmgren, J., U Söderman , U. (2002 ). Detecting and Measuring Individual Trees 700
Using an Airborne Laser Scanner.Photogrammetric Engineering & Remote Sensing 701
68(9): 925 –32. 702
Popescu, S. C. (2007). Estimating bioma ss of individual pine t rees using airborne lidar. 703
Biomass and Bioenergy, 31(9), 646 –655. http://doi.org/10.1016/j.biombioe.2007.06.022 704
Popescu, S.C., Wynne, R. (2004). Seeing the trees in the fo rest: using LiDAR and 705
multispectral data fusion with local filtering and variable window size for estimating tree 706
height. Photogrammetric Engineering and Remote Sensing 70, 589 -604. 707
Popescu, S.C., Wynne, R.H., Nelson, R.F. (2002). Estimating plot -level tre e heights with 708
LiDAR: l ocal filtering with a canopy -height based variable window size. Computers and 709
Electronics in Agriculture 37(1 -3), 71 -95. 710
Popescu, S.C., Zhao, K. (2008).A voxel -based lidar method for estimating crown base height 711
for deciduous and pin e trees. Remote Sensing of Environment 112, 767 –781. 712
Räsänen, A., Kuitunen, M., Tomppo, E., Lensu, A. (2014). Coupling high -resolution satellite 713
imagery with ALS -based canopy height model and digital elevation model in object – 714
based boreal forest habitat ty pe classification. ISPR S Journal of Photogrammetry and 715
Remote Sensing , 94, 169 –182. https://doi.org/10.1016/j.isprsjprs.2014.05.003 716
Reitberger, J., Krzystek, P., Stilla, U. (2009). Benefit of Airborne Full Waveform LIDAR for 717
3D segmentation and classification of single trees. Proceedings of the American Society 718
for Photogrammetry and Remote Sensing Annual Conference 2009, vol 2, 670 -678. 719
SenseFly.2014. sBee Extended User Manual. Available onli ne: 720
http://95.110.228.56/documentUAV/drone%20manual/[ENG]_2014_Extended_User_M 721
anual_eBee_RTK_v1_1.pdf (accessed on 9 May 2019) 722
Solberg, S. , Naes set, E., Bollands as, O. M. (2006). Single tree segmentation using airborne 723
laser scanner data in a structurally heterogeneous spruce forest. Photogrammetric 724
Engineering & Remote Sensing, 72(12), 1369 –1378. http://doi.or g/0099 -1112/06/7212 – 725
1369 726
Spracklen, B. D., Spracklen, D. V. (2019). Identifying European Old -Growth Forests using 727
Remote Sensing: A Study in the Ukrainian Carpathians. Forests, 10(2), 127. 728
http://doi.org/10. 3390/f10020127 729
Team, R.C., 2018. R: A Language and Environment for Statistical Computing [WWW 730
Document]. Vienna, Austria. 731

38
Thiel, C., Schmullius, C. (2016). Comparison of UAV photograph -based and airborne lidar – 732
based point clouds ove r forest from a forestry application perspective, International 733
Journal of Remote Sensing, DOI:10.1080/01431161.2016.122518 734
Tiede, D., Hoffmann, C., Willhauck, G. (2012). Fully integrated workflow for combining 735
object -based image analysis and LiDAR point cloud metrics for feature extraction and 736
classification improvement. ILMF International LiDAR Mapping Forum 2012, 737
(November), 6. 738
Vastaranta, M., Holopainen, M., Yu, X., Hyyppä, J., Mäkinen, A., Rasinmäki, J., Melkas, T., 739
Kaartinen, H., Hyyppä, H. (2011). E ffects of individual tree detection error sources on 740
forest management planning calculations, Remote Sensing , 3(8),1614 –1626. doi : 741
10.3390/rs3081614. 742
Vauhkonena, J., Korpelab, I., Maltamoa, M. , Tokolaa, T. (2010). Imputation of single -tree 743
attributes usin g airborne laser scanning -based height, intensity, and alpha shape metrics. 744
Remote Sensing of Environment 114 (6), 1263 –1276. 745
Vopěnka, P ., Černý, M. (2006). In Transcarpathia, Ukraine, GIS aids statistical forest 746
inventory. ArcNews 27(4). Retrived Winter 2005/2006 from 747
http://www.esri.com/new s/arcnews/winter0506articles/in -transcarpathia.html 748
Wallace, L., Lucieer, A., Malenovsk ỳ, Z., Turner, D., Vopěnka, P. (2016). Assessment of 749
forest structure using two UAV techniques: A comparison of airborne laser scanning and 750
structure from motion (SfM) p oint clouds. Forests , 7(3), 1 –16. 751
https://doi.org/10.3390/f7030062 752
Wang, Y., Hyyppä, J., Liang, X., Kaartinen, H., Yu, X., Lindberg, E., Holmgren , J., Qin, Y., 753
Mallet, C., Ferraz A., Torabzadeh, H., Morsdor f, F., Zhu, L., Liu, J. , Alho, P. (2016). 754
International Benchmarking of the Individual Tree Detection Methods for Modeling 3 -D 755
Canopy Structure for Silviculture and Forest Ecology Using Airborne Laser Scanning. 756
IEEE Transactions on Geoscience and Remote Se nsing, 54(9), 1 –17. 757
http://doi.org/10.1109/TGRS.2016.2543225 758
Weir, D. (2012). Assessment of processing different types of vegetative data using LiDAR. 759
Thesis report, University of New South Wales. 760
White, J. C., Wulder, M. A., Varhola, A., Vastaranta, M., Coops, N. C., Cook, B. D., Woods, 761
M. (2013). A best practices guide for generating forest inventory attributes from 762
airborne laser scanning data using an area -based approach. Forestry Chronicle (Vol. 89). 763
http://doi.org/10.5558/tfc2013 -132 764
Wickham, H., 2017. Package ‘tidyverse.’ Cran -R. 765
https://doi.org/10.1093/bioinformatics/btr406 766
Wittk e, S., Yu, X., Karjalainen, M., Hyyppä, J., Puttonen, E. (2019). Comparison of two – 767
dimensional multitemporal Sentinel -2 data with three -dimensio nal remote sensing data 768
sources for forest inventory parameter estimation over a boreal forest. International 769

39
Journal of Applied Earth Observation and Geoinformation, 76(December 2018), 167 – 770
178. https://doi.org/10.1016/j.jag.2018.11.009 771
Yao, W., Krzystek, P., Heurich, M. (2012). Tree species classification a nd estimation of stem 772
volume and DBH based on single tree extraction by exploiting airborne full -waveform 773
LiDAR data. Remote Sensing of Environm ent, 123, 368 –380. 774
Yu, X., Hyyppä, J., V astaranta, M., Holopainen, M., Viitala, R. (2011). Predicting individual 775
tree attributes from airborne laser point clouds based on the random forests technique. 776
ISPRS Journal of Photogrammetry and Remote Sensing, 66( 1), 28 –37. 777
http://doi.org/10.1016/j.isprsjprs.2010 .08.003 778
Zahawi, R. A., Dandois, J. P., Holl, K. D., Nadwodny, D., Reid, J. L., Ellis, E. C. (2015). 779
Using lightweight unmanned aerial vehicles to monitor tropical forest recovery. 780
Biological Conservation , 186, 287 –295. https://doi.org/10.1016/j.biocon.2015.03.031 781

Table 1. Description of the forest characteristics within all the plots .
Plot Size
(ha) Tree
species Density
(stems*ha-1) Height (m) Kraft Class
min max mean I II III IV V
41 1 Norway
spruce 208 5.2 41 29.22 57 61 48 18 24
Beech 183 4.7 39.5 25.10 16 57 57 49 4
Other
species 19 14.2 37.5 28.99 2 10 6 0 1
TOTAL 410 4.7 41 27.37 75 128 111 67 29
44 1 Norway
spruce 199 4 39.3 28.92 74 66 21 12 26
Beech 81 6.4 32.2 25.38 3 34 27 15 2
Other
species 48 7.7 37.4 23.55 4 7 20 16 1
TOTAL 328 4 39.3 27.26 81 107 68 43 29
38 1 Norway
spruce 201 4.4 35.6 25.44 86 44 28 12 31
Beech 394 4.9 34.3 22.33 34 180 133 37 10
Other
species 69 3.3 30.9 19.03 10 28 5 18 8
TOTAL 664 3.3 35.6 22.93 130 252 166 67 49
40 1 Norway
spruce 350 4 34.9 26.08 72 130 74 23 51
Beech 220 4.8 30.3 19.61 8 84 118 10
Other
species 97 4 31.8 18.71 1 10 31 39 16
TOTAL 667 4 34.9 22.87 73 148 189 180 77
Table1
Click here to download Table: Table 1.doc

Table 2. Characteristics of ALS data .
Parameter Name/Value
Sensor Airborne Lidar Scanner ‘digital full
waveform analysis ’, RIEGL Q780
Aircraft Diamond DA42 MPP aircraft
Mean flight altitude 808 m AGL
Field of view (FOV) >600
Pulse Repetition Rate (PRR ): >400 kHz
Scanning speed > 266 ,000 measurements sec-1
Point density 5–8 points m-2
Altimetric precision 20 cm
Total planimetric precision 20 cm
Table2
Click here to download Table: Table 2.doc

Table 3. Accuracy assessment of OBIA tree species’ classification.
Plot
Spruce Beech Canopy -gaps Other species
PA/UA %
PA/UA % PA/UA % PA/UA % Ovr. %
38 77.9/71.0 71.9/87.8 77.2/55.2 52.4/67.2 73.9
40 74.7/81.2 60.5/62.2 82.2/74.8 NA 75.9
41 68.9/8 1.5 78.9/84.9 86.1/66.7 NA 77.3
44 69.7/81.8 67.9/77.1 89.0/71.7 78.1/84.8 76.3
Note: UA, User’s accuracy; PA, Producer’s accuracy; Ovr, Overall accuracy
Table3
Click here to download Table: Table 3.doc

Table 4. Canopy maxima specific parameters (A, B, C, D) computed locally for tree species within each plot .
Plot Species Canopy maxima specific parameters
A B C D
41 Norway spruce 3.987 -0.0694 0.004 0
Beech 2.5072 0.0186 0.0031 0
44 Norway spruce 4.2891 -0.0681 0.0037 0
Beech 6.0432 -0.3933 0.0125 0
38 Norway spruce 3.4336 -0.0585 0.0041 0
Beech 3.6464 -0.2004 0.0086 0
40 Norway spruce 2.5422 -0.0176 0.0038 0
Beech 3.248 -0.0962 0.0073 0
Table4
Click here to download Table: Table 4.doc

Table 5. Mean height, crown diameter values of field -measured trees and matched ALS -derived mean height and crown
diameter values.
Plot Tree species Matched
trees

(m)

(m)

(m)

(m)

(m)

(m)
41 Norway spruce 96 34.37 34.92 1.41 6.44 3.15 3.69
Beech 28 31.71 32.19 1.23 5.73 3.10 2.93
44 Norway spruce 118 32.99 33.37 0.77 6.06 3.64 2.89
Beech 17 29.58 29.74 1.02 4.59 3.12 1.69
38 Norway spruce 101 31.34 30.95 0.71 5.70 3.73 2.39
Beech 67 27.91 27.96 0.86 4.85 2.99 2.17
40 Norway sp ruce 156 30.15 29 1.27 5.63 3.33 2.55
Beech 34 24.2 23.46 1.63 5.47 2.07 3.59
All
plots Norway spruce 471 31.98 31.73 1.09 5.95 3.47 2.87
Beech 146 27.98 27.93 1.17 5.13 2.81 2.67
Table5
Click here to download Table: Table 5.doc

Table 6. Linear regression for the prediction of dbh based on ALS data .
Plot Tree species Dependent variable
(field measurement) Model and significant variables :
ALS derived height (H), ALS
derived crown dimeter (CrD) R2 RMSE_
dbh
(cm)
38 Norway spruce dbh (cm) 2.003*H+1.523* CrD -21.053 0.62 4.8
Beech 0.960*H+1.490*CrD -1.992 0.17 6.2
41,44,40 Norway spruce dbh (cm) 1.675*H+1.504* CrD -9.345 0.56 6.1
Beech 0.624*H +2.183* CrD+6.597 0.42 5.6

Table6
Click here to download Table: Table 6.doc

Table 7. Mean dbh measured in the field and mean predicted dbh based on ALS data for the same trees and T -paired,
Wilcoxon test statistical significance s.
Plot Tree
species Number
of
matched
trees
Fielddbh(cm)
ALSdbh(cm) Freed om
degree Tvalue,
0.05 pvalue Wilcoxon
Zvalue Wilcoxon
pvalue
38 Norway
spruce 201 46.616 46.619 200 -0.002 0.998
beech 66 28.912 28.912 -0.687 0.492
41,44,
40 Norway
spruce 372 49.25 49.25 -0.288 0.774
beech 79 29.82 28.83 -1.676 0.094
Table7
Click here to download Table: Table 7.doc

Table 8. Mean volume of trees calculated based on field measurements and mean estimated volume based on ALS data —
Wilcoxon test and statistical significance.
Plot
Tree
species Number of
matched
trees
Fieldv(m3)
ALSv (m
3) Wilcoxon
Zvalue Wilcoxon
pvalue
38 Norway
spruce 100 2.29 2.24 -.087 0.930
beech 66 0.97 0.92 0.457 0.648
41,44,40 Norway
spruce 372 2.59 2.56 -0.550 0.582
beech 79 1.07 0.97 1.933 0.053
Table8
Click here to download Table: Table 8.doc

Table 9. Percentage error of estimated total volume based on ALS data from total field reference volume .
Plot Tree
species Total
ground
reference
volume
(m3) ALS
estimated
volume of
matched
trees
(m3)
Field reference
volume of ALS –
matched trees
(m3) Percentage of the
total ALS -estimated
volume
compared to
total field
reference
volume
(%) RMSE of
volume
(m3)
0 1 2 3 4 5

6
41 Norway
spruce 438.1 317.7 304.4 73 0.6
Beech 180.0 38.6 39.5 21 0.5
44 Norway
spruce 433.0 342.7 342.6 79 0.6
Beech 79.6 19.0 22.6 24 0.5
38 Norway
spruce 298.7 224.0 229.2 75 0.4
Beech 210.3 61.4 64.6 29 0.4
40 Norway
spruce 504.0 292.8 316.6 58 0.4
Beech 88.1 17.2 23.0 20 0.2
Table9
Click here to download Table: Table 9.doc

Figure1
Click here to download high resolution image

Figure2
Click here to download high resolution image

Figure3
Click here to download high resolution image

Figure4a
Click here to download high resolution image

Figure4b
Click here to download high resolution image

Figure4c
Click here to download high resolution image

Figure4d
Click here to download high resolution image

Figure5
Click here to download high resolution image

Figure5b
Click here to download high resolution image

Figure6a
Click here to download high resolution image

Figure6b
Click here to download high resolution image

Figure6c
Click here to download high resolution image

Figure6d
Click here to download high resolution image

Figure7a
Click here to download high resolution image

Figure7b
Click here to download high resolution image

Figure7c
Click here to download high resolution image

Figure7d
Click here to download high resolution image

Figure8a
Click here to download high resolution image

Figure8b
Click here to download high resolution image

Figure8c
Click here to download high resolution image

Figure8d
Click here to download high resolution image

Figure9a
Click here to download high resolution image

Figure9b
Click here to download high resolution image

Figure10a
Click here to download high resolution image

Figure10b
Click here to download high resolution image

Figure11a
Click here to download high resolution image

Figure11b
Click here to download high resolution image