You are on page 1of 13

Scandinavian Journal of Forest Research

ISSN: 0282-7581 (Print) 1651-1891 (Online) Journal homepage: http://www.tandfonline.com/loi/sfor20

Assessing 3D point clouds from aerial photographs


for species-specific forest inventories

Stefano Puliti, Terje Gobakken, Hans Ole Ørka & Erik Næsset

To cite this article: Stefano Puliti, Terje Gobakken, Hans Ole Ørka & Erik Næsset (2017)
Assessing 3D point clouds from aerial photographs for species-specific forest inventories,
Scandinavian Journal of Forest Research, 32:1, 68-79, DOI: 10.1080/02827581.2016.1186727

To link to this article: https://doi.org/10.1080/02827581.2016.1186727

Accepted author version posted online: 16


May 2016.
Published online: 21 May 2016.

Submit your article to this journal

Article views: 464

View Crossmark data

Citing articles: 17 View citing articles

Full Terms & Conditions of access and use can be found at


http://www.tandfonline.com/action/journalInformation?journalCode=sfor20
SCANDINAVIAN JOURNAL OF FOREST RESEARCH, 2017
VOL. 32, NO. 1, 68–79
http://dx.doi.org/10.1080/02827581.2016.1186727

Assessing 3D point clouds from aerial photographs for species-specific forest


inventories
Stefano Puliti, Terje Gobakken, Hans Ole Ørka and Erik Næsset
Department of Ecology and Natural Resource Management, Norwegian University of Life Sciences, Ås, Norway

ABSTRACT ARTICLE HISTORY


In this study we assessed the potential of using photogrammetric data for species-specific forest Received 4 December 2015
inventories. The method is based on a combination of Dirichlet and ordinary linear regression Accepted 1 May 2016
models. This approach was used to predict species proportions, main tree species, total, and
KEYWORDS
species-specific volume. Structural and spectral variables were used as predictors. The models were Species-specific forest
validated using 63 independent validation stands. The results from airborne laser scanning (ALS) management inventory;
data combined with spectral data and photogrammetric data obtained using aerial imagery with photogrammetry; airborne
different forward overlaps of 80% and 60% were compared. The best photogrammetry-based laser scanning; Dirichlet
models predicted species proportions with a relative root mean square error (RMSE) of 21.4%, regression; species
classified dominant species with 79% accuracy, predicted total volume with relative RMSE of 13.4%, proportions
and predicted species-specific volume with relative RMSE of 36.6%, 46.5%, and 84.9% for spruce,
pine, and deciduous species, respectively. The results were similar for the three point cloud
datasets obtained from aerial imagery and ALS and the accuracies of the predictions were
comparable to methods used in operational FMI. The study highlights the effectiveness of forest
inventories carried out using photogrammetric data, which – differently from ALS, can include
species-specific information without relying on multiple data sources.

Introduction used to make predictions on a wall-to-wall basis. Except for


Bohlin et al. (2012b), the focus of the previous studies has
Since around year 2010, there has been an increasing interest
been exclusively on height and density variables extracted
in the use of three-dimensional (3D) point clouds derived
from digital aerial photogrammetry to model biophysical from the 3D point clouds and on the prediction of biophysical
forest properties because of the smaller acquisition costs forest properties like timber volume, basal area, and stem
compared to airborne laser scanning (ALS) data. The potential number. Bohlin et al. (2012b) first suggested the possibility
for use of photogrammetric point clouds to model biophysical to use also the spectral information from digital aerial
forest properties such as timber volume has been investi- images to predict species-specific information, which rep-
gated in several recent studies (e.g. Bohlin et al. 2012a; resents an important requirement for FMIs (Packalén &
Järnstedt et al. 2012; Nurminen et al. 2013; Straub et al. Maltamo 2007; Packalén et al. 2009; Ørka et al. 2013;
2013; Vastaranta et al. 2013; Gobakken et al. 2014; Rahlf Korpela et al. 2014). One of the advantages of using photo-
et al. 2014) by comparing the accuracy obtained from photo- grammetric 3D point clouds compared to ALS is the possibility
grammetric and ALS point clouds. Despite intrinsic differences to directly provide multispectral and structural information
between the studies, the differences for timber volume in without having to rely on the fusion of multiple data
relative root mean square error (RMSE; as percentage of sources. Thus, photogrammetry-based species-specific FMIs
ground reference value) between photogrammetric and are not subject to technical, economical, and weather limit-
ALS-based predictions are reported to be 0.1–12.0% at plot ations related to the separate acquisitions of ALS and multi-
level (250–500 m2) and 0.5–5.8% at stand level. These pub- spectral data (Korpela et al. 2011; Vauhkonen et al. 2014).
lished results verified the possibility to effectively implement Spectral and textural information have been widely used to
photogrammetry-based forest management inventories differentiate between different land use classes (Haralick et al.
(FMIs) as an alternative to ALS. 1973; Franklin et al. 2000; Waser et al. 2011). In forest remote
Most published studies using photogrammetric point sensing, digital imagery has been used to discriminate
clouds for FMI have followed the area-based approach between different tree species. The transition from film to
(ABA). For each sample plot several variables describing the digital imaging enabled the development of methods to
vertical and horizontal distribution of the points are extracted delineate and classify single trees with different degrees of
from the normalized point cloud and used to fit models with automation (Haara & Haarala 2002; Erikson 2004; Bohlin
ground reference values of the forest biophysical properties et al. 2006; Warner et al. 2006). The proposed methods were
as dependent variables. The estimated models are then promising and were later refined using a fusion of ALS and

CONTACT Stefano Puliti stefano.puliti@nmbu.no Department of Ecology and Natural Resource Management, Norwegian University of Life Sciences,
P. O. Box 5003, NO-1432 Ås, Norway
© 2016 Informa UK Limited, trading as Taylor & Francis Group
SCANDINAVIAN JOURNAL OF FOREST RESEARCH 69

multispectral imagery. The addition of high resolution canopy spruce and Lodgepole pine (Donoghue et al. 2007).
height data to the multispectral data was proven to further However, beta regression is a more appropriate alternative
improve single-tree delineation and classification (Leckie to model such proportions. Beta regression has been used
et al. 2003; Holmgren et al. 2008; Ørka et al. 2012). Due to a to model crown coverage from 3D point clouds (Korhonen
great deal of interest in operational application of ABA com- et al. 2007), understory vegetation coverage (Wing et al.
pared to single-tree approaches, in Finland, an ABA method 2012), and species compositions (Vihervaara et al. 2015).
was developed to model species-specific biophysical forest Nevertheless, when more than two species are present,
properties based on a combination of aerial imagery and Dirichlet models (Maier 2014) are a statistically sound
ALS data (Packalén & Maltamo 2007; Packalén et al. 2009). method which effectively handle simultaneous occurrence
The findings from Packalén and Maltamo (2007) proved that of a variety of species in a given landscape (Hijazi & Jernigan
the achieved results were as accurate as conventional stand 2007; Maier 2014). In addition, this approach can potentially
level FMIs in Finland. The method is currently applied opera- represent an alternative to non-parametric methods in
tionally on a large scale (Maltamo et al. 2011; Maltamo et al. many inventories where the number of sample plots is small.
2014a). Regardless the advancements in radiometric and geo- Tree species distribution has shown a notable impact on
metric properties of digital images, there still are several the relationship between 3D remotely sensed data and
factors posing challenges in their operational use for quanti- forest biophysical properties (Næsset 2005; Ørka et al. 2010).
tative analyses of forests (Honkavaara et al. 2009; Korpela When using stand level density and timber volume models
et al. 2011). Difficulties in the elimination of atmospheric developed in conifer-dominated forests to make predictions
effects, issues of heterogeneity of within-species reflectance in pure deciduous stands, Næsset (2004b) found a mean
properties and directional reflectance anisotropy can seriously difference (MD) between predictions and ground reference
affect species classification (Korpela et al. 2011). of up to 90%. Such overestimation was partially explained
In Norway, operational FMIs based on ALS rely on a first by the different crown shapes and composition (i.e. leaves
phase of manual image interpretation during which pro- versus needles and foliage versus branches) between conifers
portions of tree species and their age are determined and and deciduous species. Ørka et al. (2010) also found that when
used to stratify the forest area for subsequent estimation using a single-tree approach, the inclusion of tree species
(Næsset 2004a). This phase is time consuming and subjective improved tree height and stem diameter prediction accuracy.
(Eid & Næsset 1998; Ørka et al. 2013; Maltamo et al. 2014b; Thus, as suggested by previous research, species information
Næsset 2014) and results in large costs and potential bias of should be taken into account when modeling such properties.
the estimators if calibration is not carried out properly. The It is especially relevant to highlight the importance of the pro-
development of automated and reliable methods to derive portion of deciduous trees on plots and in stands, as decid-
species-specific forest biophysical properties would therefore uous trees differ substantially from conifers in crown shape,
be beneficial to the Norwegian forest sector (Næsset 2014). resulting in different relationships between 3D point clouds
The output from species-specific inventories varies by and biophysical properties. In this regard, the inclusion of pre-
country and management regime. In Finland, species-specific dicted species proportions as explanatory variables when
values are required for all forest biophysical properties, while modeling forest biophysical properties is expected to
in Norway species proportions in relation to timber volume improve the accuracy of the developed models.
are the variables of interest. Main tree species, defined as The flight settings used during the image acquisition affect
the tree species with the largest basal area proportion the quality of the 3D (Bohlin et al. 2012a) and the spectral
(Mora et al. 2010; Ørka et al. 2013) or volume proportion (Pack- information extracted from the imagery. The percentage of
alén et al. 2009) is also a required property from FMIs. longitudinal overlap represents one of the parameters that
Previous species-specific studies of FMI have often relied can increase the quality of the abovementioned image-
on non-parametric imputations to predict plot and stand derived products without incurring major increase in costs.
level species-specific forest biophysical properties (Packalén The quality of the photogrammetric point clouds is potentially
and Maltamo 2007; Packalén et al. 2009; Niska et al. 2010). A improved when image overlap is increased due to the
fairly large sample of plots is then used. Studies relying on reduced occlusions (Bohlin et al. 2012a; Nurminen et al.
smaller number of sample plots, as used in Norway, have 2013). Thus, resulting in point clouds with larger height vari-
also applied non-parametric methods such as regression ations, which ultimately reflects on the amount of information
trees (Ørka et al. 2013). As mentioned above, one of the fun- provided by the height and density variables. Nevertheless,
damental properties of interest in many countries is the previous studies investigating the effect of varying image
species proportions. Additionally, from species proportions it overlap on the quality of the prediction of forest biophysical
is possible to derive species-specific values of the forest bio- properties using ABA (Bohlin et al. 2012a; Nurminen et al.
physical properties of interest, such as for example timber 2013), concluded that there was no significant increase in
volume, as well as the classification of the stand dominant the accuracy of the predictions with greater longitudinal over-
species. Species proportions are compositional data, where laps. The increase in accuracy of the prediction of species-
the different components are represented by the proportion specific forest biophysical properties may be greater than
of each species in relation to the total of a certain forest bio- for total values (non-species-specific properties) due to the
physical property (e.g. timber volume and basal area). Statisti- fact that larger overlaps produce more homogeneous point
cal models to model such compositional data exist. Logistics clouds, therefore increasing the number of point observations
regression has been used to model proportions of Sitka available for each tree.
70 S. PULITI ET AL.

The present study applied a novel method based on a used to calibrate the models and was composed of 151 circu-
combination of Dirichlet regression and linear regression lar plots of size 400 m2 each (radius = 11.2 m). The sample
using photogrammetric data and a pre-existing high resol- plots were systematically distributed throughout the study
ution digital terrain model. Three different 3D point cloud area according to a regular grid. The stand validation
datasets were compared. To evaluate the performance of dataset was used to evaluate the accuracy of the developed
the method under a two of imagery acquisition scenarios, models. It was composed of 63 stands and a total of 744
UltraCam XP multispectral imageries with different forward plots were systematically distributed within the stands with
overlap of 80% (IM80) and 60% (IM60) were compared. The separate sample grids in each stand. The original design
IM60 represented the business-as-usual scenario for aerial aimed at obtaining 15 plots per stand, however, the number
photography in various public programs in Norway, whereas was reduced proportionally if parts of the stand were har-
IM80 characterized an alternative to improve the quality of vested. The actual number was on average 12 with a
the end results. Since ALS is used as a standard technology minimum of 3 and a maximum of 16. Plot size (i.e. 125 m2
in current FMIs in Norway (Næsset 2014), an ALS dataset or 250 m2) was defined according to forest development
was used as benchmark to determine the applicability of class. Further details about the varying sample plot sizes in
photogrammetric species-specific FMI. the stand validation dataset can be found in Gobakken
The objectives of this study were to assess the accuracy of et al. (2014). The portion of measured stand area relative to
predicting (1) a number of relevant biophysical forest proper- the total stand area ranged from 5.5% to 67.5% with a
ties (2) at different spatial scales (plot and stand level) and (3) mean of 37.3%.
using different 3D data sources (ALS, IM80, and IM60). The fol- Below the common measurements and calculations done
lowing properties were assessed: for the two datasets are described. Diameter at breast
height (DBH) was measured for all trees with DBH ≥4 cm in
. species proportions; circular plots. Height sample trees were selected based on
. dominant species; probability proportional to stem basal area and measured
. total volume; using a Vertex hypsometer. The volume of each height
. species-specific volume. sample tree was predicted using species-specific volume
models (Braastad 1966; Brantseg 1967; Vestjordet 1967)
using DBH and measured height as the predictor variables.
Materials Subsequently, the volume of the same sample trees was com-
puted using the same models but with predicted height. For
Study area
tally trees (i.e. trees with no height measurements), height
The study was conducted in a 852.6 ha boreal forest area in was predicted using height-DBH models by Fitje and Vestjor-
Våler municipality (59°30′ N, 10°55′ E, 70–120 m a.s.l.) located det (1977) and by Vestjordet (1968). The volume of each tally
in south-eastern Norway. The dominant tree species is tree was predicted using species-specific volume models
Norway spruce (Picea abies (L.) Karst.) followed by Scots pine (Braastad 1966; Brantseg 1967; Vestjordet 1967) using DBH
(Pinus sylvestris L.) representing 51% and 36% of the timber and predicted height as predictor variables. The volume esti-
stock, respectively. The remaining proportion (13%) is decid- mates of tally trees were then adjusted by the plot and
uous trees dominated by Birch (Betula pubescens Ehrh.). The species-specific ratio of the mean volume estimate for
forest area is actively managed aimed at the timber pro- height sample trees with predicted heights and the mean
duction. Silvicultural treatments are applied in accordance volume estimate for height sample trees with measured
with standards in boreal forest ecosystems across the Scandi- heights. Species-specific ground reference volume (VSS) was
navian Peninsula. Clear-felling or shelterwood cutting are computed as the sum of the volume of each individual tree
applied at the end of the rotation depending on the site fer- by species. Total volume (Vtot) was calculated as the sum
tility with the latter being more common on poorer sites over all trees. Spruce, pine, and deciduous proportions (S% ,
dominated by pine. The regeneration is achieved by planta- P% , and D% ) were computed as the proportion of each
tion in case of clear-felling and by natural regeneration after species to the total volume.
selective logging. Further details about the study area can Stand level Vtot and VSS were computed as the area-
be found in Næsset (2002). weighted mean of the plots within each stand. A summary
of the ground reference values in the sample plot dataset
and stand validation dataset is presented in Table 1.
Field data collection
The field campaign was conducted during summer 2010 and
Remotely sensed data
finished in spring 2011. The present study considered all
forests with mean height larger than approximately 7 m, The remotely sensed data were collected under leaf-on con-
according to the Norwegian forest classification, as these rep- ditions during July 2010 with a Piper Navajo fixed-wing air-
resent the forests that are operationally inventoried using ALS craft. The ALS data were acquired on 2 July 2010 with an
in Norway. The field measurements were divided into two dis- Optech ALTM-Gemini laser scanner sensor and result in an
tinct datasets, namely: (1) sample plot dataset from a systema- average pulse density was 7.45 points m−2 (Table 2). ALS
tic survey across the entire area and (2) stand validation data pre-processing was carried out by the contractor (Blom
dataset from selected stands. The sample plot dataset was Geomatics, Norway). This included computation of
SCANDINAVIAN JOURNAL OF FOREST RESEARCH 71

Table 1. Summary statistics of total volume (m3 ha−1), species-specific volume flight strips (i.e. side overlap). Aircraft location and orientation
(m3 ha−1) for spruce, pine, and deciduous, and the species proportion (%) of the during image acquisition were recorded using a GPS and an
total volume (Vtot) for the sample plot dataset and the stand validation dataset.
inertial navigation system.
Standard Proportion of
Minimum Maximum Mean deviation Vtot
Sample plot dataset (n = 151)
Spruce 0.78 794.51 146.58 145.66 51 Methods
Pine 0.00 274.81 70.78 63.63 36
Deciduous 0.00 273.23 30.95 41.19 13 Photogrammetric processing
Total 57.23 800.78 248.31 126.13 100
Stand validation dataset (n = 63) The photogrammetric processing was carried out for two
Spruce 7.556 526.73 127.26 116.24 48 different sets of imagery (i.e. IM80 and IM60). The IM60
Pine 0.00 208.51 79.13 46.71 40
Deciduous 0.29 129.14 27.05 27.81 12 dataset was created by removing every other image from
Total 111.33 545.46 233.50 101.46 100 the IM80 dataset. When removing every other image in the
strips that were acquired with 80% forward overlap the dis-
tance between the center points of consecutive images is
planimetric coordinates and ellipsoid height values, and 40% of the image length, thus the 60% forward overlap is
classification of the point cloud into ground and non- obtained. Photogrammetric 3D point clouds were produced
ground echoes according to the proprietary algorithm from the panchromatic imagery using the proprietary soft-
implemented in Terrascan software (Terrasolid 2012). A trian- ware Agisoft PhotoScan version 10.1 (Agisoft 2014) as it has
gulated irregular network (TIN) surface was then created by been proven to be suitable for forest inventory purposes
linear interpolation from the ground-classified points. (Puliti et al. 2015). The panchromatic image was used since
Multispectral data were acquired on 24 July 2010 with a it provided the highest spatial resolution. Provided the
Vexcel UltraCam Xp imaging sensor. The acquisition lasted for camera interior and exterior orientation, the algorithm
approximately 7 minutes and was carried in two adjacent implemented searches and matches common features
flight strips. The camera system consists of four panchromatic points on contiguous images. The main output of this first
and four spectral camera heads. The resulting level-2 images stage is the estimated and optimized camera internal and
(n = 35) consisted in one full resolution panchromatic image external parameters. Subsequently, point cloud densification
(11,310 pixels × 17,310 pixels) and one multichannel color was performed using Photoscan proprietary multiview
image at color resolution (3770 pixels × 5770 pixels). The latter stereo-reconstruction algorithms. The densification was com-
one was composed of four bands representing the red, green, puted using “high” and “medium” quality settings (Agisoft
blue, and near infra-red (RGBI) wavelengths. The images had a 2014) for IM80 and IM60, respectively. These settings were
spectral resolution of 16 bit and no radiometric correction was defined using a trial-and-error approach, as the ones produ-
performed by the data vendor. The ground sampling distance cing the most homogeneously distributed point cloud. Point
was of 0.17 m for the panchromatic images and 0.51 m for the cloud filtering was carried out using “mild” settings for IM80
RGBI images. The acquisition was performed using an 80% and IM60, as this has shown to be the most suitable setting
overlap between consecutive images within a flight strip (i.e. for forestry purposes since it ensures noise reduction while
forward overlap) and 30% overlap between images in adjacent maintaining the largest height variations between points
(Puliti et al. 2015). Finally, the reconstructed geometries (i.e.
mesh) from the dense point cloud were used to generate pan-
Table 2. Summary of technical specifications of the acquisitions and the sensors. chromatic orthophotos. Thus, the outputs exported from
Technical specifications Photoscan consisted of (1) two panchromatic orthophotos
Platform Piper Navajo fixed-wing aircraft and (2) two dense point clouds representing IM80 and IM60.
Optech ALTM- The ground sampling distance was set to 0.17 m for both
Sensor Vexcel UltraCam Xp Gemini datasets with different overlaps. The point densities of the
Mean flying speed (m s−1) 80 80 resulting point clouds were 12.0 and 2.3 points m−2 for the
Flying altitude above ground (m) 2850 900
Number of flight lines 2 18 IM80 and IM60 datasets, respectively. The lower forward
Side overlap (%) 30 55 overlap resulted in a reduced number of matched points in
Forward overlap (%) 80 – canopy gaps and therefore in lower height variations within
Scan frequency (Hz) – 55
Half-scan angle (degrees) – 13.8 the point cloud. The point cloud derived from IM80 was
Swath width (m) 2940 440 later thinned to 50% of the full dataset to reduce time in
Pulse repetition frequency (kHz) – 100 the subsequent processing as previous research has shown
Max. number of echoes per pulse – 4
Pulse density (m−2) – 7.4 that a reduced point density reduces the accuracy of ABA
Spatial resolution (cm)a,b Panchromatic: 17a 19b FMI only marginally when the resulting density is greater
RGBI: 51a than, say, 0.5 to 0.25 points m−2 (Gobakken & Næsset 2008).
Spectral resolution (nm) 410–570 1064
480–630 Another reason for the thinning of the IM80 point cloud
580–700 was to obtain a point cloud with resolution (i.e. average
690–1000 point spacing) matching the ground sampling distance of
a
Ground sampling distance. the RGBI imagery (i.e. 0.51 m). The resulting point cloud had
b
Footprint diameter (cm) computed after (Baltsavias 1999) based on mean
acquisition settings. a point density of 6 points m−2 and an average point
spacing of approximately 0.41 m.
72 S. PULITI ET AL.

Point cloud pre-processing panchromatic images using the “glcm” package (Zvoleff
2015) available in R project (R CoreTeam 2015) as this has
The absolute heights of the three point cloud datasets (ALS,
been previously used in other forest inventory studies
IM80, and IM60) were normalized by subtracting the terrain
(Hansen et al. 2015). These summed up to a total of eight vari-
height from the ALS TIN surface from each point’s height
ables for each dataset (HARmean, HARvariance, HARhomogeneity,
value. Finally, the multispectral information was assigned to
HARcontrast, HARdissimilarity, HARentropy, HARsecond moment, and
each 3D point from the non-orthorectified RGBI images by
HARcorrelation) computed with a window size and shift selected
using a rotation matrix (Mikhail et al. 2001). The all set of
using a trial-and-error approach of 11 × 11 and 3 × 3, respect-
images (n = 32) was used for the ALS and IM80 datasets in
ively. For the ALS dataset the panchromatic orthophoto
order to have a comparison between best case scenarios.
resulting from IM80 photogrammetric processing was used
The reduced set of images (n = 16) was used for the IM60
to extract the Haralick variables. In total 43 variables were
point cloud in order to produce realistic results under a
extracted.
business-as-usual scenario. The interior and exterior orien-
tation parameters were derived from the camera calibration
report and from the aerial triangulation, respectively. These Model training
were used in addition to the x, y, and z coordinates of each
3D point to co-register the point clouds with the non-orthor- In the present study a two-step approach was used to predict
ectified imagery. The digital number of the pixel relative to species-specific biophysical forest properties. In the first step,
the image that had the center-point closest to the position two models were fitted from the sample plot dataset for ALS,
of each 3D point was assigned. For ALS, the pixel values IM60, and IM80. One model was developed for predicting S% ,
were assigned only to the first return echoes since these rep- P% , and D% and another one for Vtot. In the second step, the
resent the canopy surface (Packalén et al. 2009; Ørka et al. selected models were used to predict the variables of interest
2012). The multispectral data were radiometrically corrected at stand level. Furthermore, the dominant species was classi-
by dividing the digital number of each band (i.e. red, green, fied as the one with the largest predicted species proportion
blue, NIR) by the sum of the digital number values of all of (S% , P  
% , or D% ) and predicted species-specific volume (VSS )
  
was computed as the product of S% , P% , or D% and predicted
the bands corresponding to the same point. This method
total volume (V  tot ).
was used as it has proved to improve the accuracy in
species classification by reducing the differences in radiance Tree species proportions were modeled using structural
among different images (Yu et al. 1999; Wu 2004; Dalponte and spectral variables by means of Dirichlet regression,
et al. 2013; Dalponte et al. 2014). implemented in the R package DirichletReg (Henningsen &
Toomet 2011; Maier 2014). Dirichlet regression allows to
model compositional data, which by definition are non-nega-
Variable extraction tive, and sum up to unit (Hijazi & Jernigan 2007). The par-
ameters in the Dirichlet model are estimated by maximum
First, structural variables were extracted from the three point likelihood (Hijazi & Jernigan 2007). The matrix describing the
cloud datasets (ALS, IM80, and IM60). These included height volume proportions of spruce, pine, and deciduous (S% , P %,
percentiles (p10, … , p95, p100), height standard deviation or D % ) for each sample plot, was used as dependent variable
(hsd), and density variables (Næsset 2004b). The latter ones in the modeling. However, due to mathematical reasons the
were calculated for equally spaced vertical layers, defined as vector has to include values between 0 and 1. When the pro-
tenths of the distance between the 95th percentile and the portions were smaller than 1.49e-08 or larger than 1–1.49e-08
lowest canopy height (i.e. 2 m). The densities were computed a transformation of all the values in the matrix was carried out
as the proportion of points above the 1st, … ,9th (d0, … ,d9) with the DR_data function (Maier 2014). The transformation
fraction to the total number of points. A 2-m height threshold relies on the formula (Equation (1)) proposed by Smithson
was applied to separate the canopy component from the low and Verkuilen (2006):
vegetation (Nilsson 1996; Næsset 2002). The total number of
y(n − 1) + (1/d)
height and density variables for each dataset was 21. For y∗ = , (1)
the ALS dataset all echoes were used to compute the struc- n
tural variables. where y ∗ is the transformed value of the proportion, y is the
Second, spectral variables from the digital number values original proportion, n is the number of observations, and d
were extracted. The variables extracted from the image data is the number of modeled species.
were divided into tonal and textural variables. The tonal vari- The Dirichlet regression models simultaneously all the
ables were extracted from the colored point cloud and species proportions. To select a suitable set of variables to
included mean band values (Rm, Gm, Bm, and NIRm), bands’ describe all the species in the Dirichlet regression a variable
standard deviation (Rsd, Gsd, Bsd, and NIRsd), and band ratios selection method based on ordinary linear regression was
(Rred/green, Rred/blue, Rred/NIR, Rgreen/blue, Rgreen/NIR, and Rblue/NIR). applied to each species separately and then the selected vari-
Also for tonal variables a 2-m threshold was applied to ables were later merged into a single pool. Structural and spec-
exclude the color information of the points representing the tral variables, including their logarithmic and square root
ground or lower vegetation. The total number of tonal vari- transformation, and the interaction between height and
ables for each dataset was 14. The textural variables (Haralick density variables were used in the variables selection. For
et al. 1973) were extracted from the orthorectified each species proportion, ordinary linear regression models
SCANDINAVIAN JOURNAL OF FOREST RESEARCH 73

were fitted and using a branch-and-bound search was carried out based on MD, RMSE, relative MD, and relative
implemented in the R-package “leaps” (Lumley & Miller 2009) RMSE:
the best subset was selected. The search was restricted to n
i=1 (ŷ i − yi )
models with maximum five variables as this was found to be MD = , (2)

n
a suitable number of explanatory variables using a trial-and- n
i=1 (yi − ŷ i )
2
error approach. The model selection was performed using RMSE = , (3)
n
the Bayesian information criterion. Furthermore, the selection
procedure was penalized for collinearity using the variance where n is the sample size (i.e. number of plots or stands), yi
inflation factor. The subset with one variable less was iteratively is the ground reference value for the biophysical property of
selected if any of the variables in the current subset had a var- interest (i.e. S% , P% , D% , Vtot , and VSS ) for observation i, and
iance inflation factor ≥5. The variable selection resulted in a ŷi is the predicted value referred to the same unit. The V tot
maximum of five variables per species. Thus, for Dirichlet predicted values were back-transformed to arithmetic scale
model, which handles all species simultaneously, the selected using exponential function and multiplied by the correction
variables for each species were aggregated to a single set of ratio proposed by Snowdon (1991). The correction ratio was
maximum fifteen variables (i.e. sum of maximum five variables computed as the ratio of the mean ground reference values
per specie). Further reduction of the number of explanatory to the mean of the back-transformed predicted values. The
variables was performed by removing variables that were not relative MD and relative RMSE were calculated as the per-
significant (p > .1) in the Dirichlet regression. centage of the average value of the modeled forest attri-
Sample plot Vtot was fitted using a linear model with bute of interest. The evaluation of the accuracy of the
logarithmic transformations of the dependent variable, as total volume models included also an analysis of the good-
this was found to be suitable for the prediction of stand ness of fit by including the adjusted R 2 as measure. To
level forest biophysical properties (Næsset 2002). Indepen- determine the quality of the information about the main
dent variables were selected from structural variables and tree species for the validation units (i.e. plots or stands),
cross validated S% , P 
% , and D% using the variable selection the accuracy of the classification of the dominant species
method. The leave one out cross validation (CV) was carried in terms of timber volume was analyzed. Confusion matrices
out at plot level by leaving out each ith plot iteratively, the were constructed and used to determine overall and
model was fitted using the remaining plots, the estimated species-specific accuracy. The user’s accuracy corresponds
parameters were used to predict the variable of interest to the commission error calculated as the fraction of cor-
(S% , P  
% , D% , and Vtot ) for the ith plot. The procedure was rectly classified observations for each class with regard to
used both to predict species proportions and later to validate the number of observations claimed to be in that same
the models at plot level. class by the classification. The producer’s accuracy corre-
sponds to the omission error and is equal to the ratio of
the number of correctly classified observations for one
Validation
species class and the total number of observations labeled
Results were evaluated at plot level by means of leave one out with the same class. The overall accuracy is the ratio of
CV using the sample plot dataset and at stand level using the the number of correctly classified observations relative to
independent stand validation dataset. The plot level vali- the total number of observations. The kappa is a more
dation was carried out using the procedure described robust measure of the total accuracy compared to the
above. The stand level validation was performed by first sub- overall accuracy since it accounts for agreement occurring
dividing the stands into quadratic prediction units (grid cells) by chance. The significance of the differences of stand
with the same area as the sample plots (400 m2). For each grid level predictions (S% , P 
% , D% , and V tot ) between each
cell, the remotely sensed data were extracted and the struc- point cloud dataset was tested using a pairwise t-test
tural and spectral variables computed. Predicted species pro- using the Bonferroni adjustment (Miller 1981) and with
portions (S% , P 
% , and D% ) were computed based on the fitted 95% confidence level. Additionally, for V  tot to reduce some
Dirichlet models. Stand level S% , P  
% , D% , and Vtot values were of the uncertainty concerning the degree to which the esti-
calculated as the weighted mean of the cells within each mate deviates from the true value the confidence intervals
stand. For S% , P  
% , and D% the grid Vtot was used as a of the true population mean were computed and the pre-
weight, whereas grid area was used for stand V tot . To diminish diction of the mean predicted stand value was compared
the effect of prediction on sliver cells occurring along the to this interval.
stand borders due to the intersection of the regular 20 m ×
20 m grid and the irregular stand shape, sliver cells were dis- Results
carded from the prediction when their area was <100 m2.
Species proportions
From the predicted species proportions and the predicted
total volume, dominant species, and species-specific timber For each point cloud dataset (i.e. ALS, IM80, IM60), a model
volume were computed. was developed using a maximum of twelve independent vari-
The evaluation of the results consisted in assessing the ables (Table 3). The selected spectral variables in the models
accuracy of species-specific volume distribution in terms included the standard deviation of the radiometrically cor-
of S% , P  
% , D% , dominant species classification, Vtot and VSS .
 rected band values, band ratios, band mean values, and tex-
Except for the dominant species classification, the evaluation tural features (i.e. mean, variance, and entropy). For each
74 S. PULITI ET AL.

Table 3. Variables included in the Dirichlet regression to model species Table 4. RMSE between predicted and ground reference species proportions at
proportion (S% , P% , and D
% ), including transformations and interactions. The plot and stand level. RMSE in percentage of the ground reference value is
results are presented for each of the studied point cloud datasets: airborne presented in parentheses. The results are presented for each of the studied
laser scanning (ALS), UltraCam XP imageries with 80% forward overlap (IM80) point cloud datasets: airborne laser scanning (ALS), UltraCam XP imageries
and 60% forward overlap (IM60). with 80% forward overlap (IM80) and 60% forward overlap (IM60).
ALS IM80 IM60 Dataset Validationa Spruce Pine Deciduous
log(Rsd) p10 Bsd ALS Plot 20.5 (40.3) 19.5 (53.7) 11.6 (90.2)
log(NIRsd) d0 p60 Stand 16.9 (35.2) 17.4 (43.1) 9.2 (79.5)
p95 Bsd log(d7) IM80 Plot 20.91 (41.1) 19.78 (54.6) 11.7 (90.6)
d3 HARentropy log(Rm) Stand 16.2 (33.9) 16.8 (41.6) 8.7 (74.7)
log(Bsd) p100 log(NIRsd) IM60 Plot 21.4 (42.0) 20.3 (56.2) 11.8 (91.7)
log(HARvariance) log(HARmean) sqrt(Gsd) Stand 16.8 (35.0) 17.5 (43.3) 9.3 (80.1)
Gsd Gsd sqrt(Bm) a
Plot level leave one out cross validation or stand level independent validation
sqrt(Bm) log(Rm) p60 * log(d7) dataset.
p95* d3 log(Bm)
p10 * d0
p100* d0
log(Rm)/log(Gm) Dominant species classification
Notes: p10, p60, p95, p100 : percentiles of points height for 10, 60, and 95, 100%
(m); d0, d3, d7,: canopy density corresponding to the proportions of points The plot level cross validated overall accuracy of the classifi-
above fraction no. 1, 4, 8 to total number of points; Rm, Gm, Bm, Rm: mean cation of the dominant species (Table 5) revealed that IM80
band values for red, green, and blue bands; Rsd, Gsd, Bsd, NIRsd: standard devi- obtained the largest overall accuracy (79%) compared to
ation for red, green, blue, and NIR bands; HARvariance, HARentropy: Haralick var-
iance and entropy. IM60 (75%) and ALS (75%). The species classification followed
log(), sqrt(): logarithmic and square root transformations. a similar pattern as the species proportion RMSE, with a
pi * di: interaction between height percentiles and density variables. greater accuracy in the classification of spruce-dominated
plots (84–88%) compared to pine (73–77%) and deciduous
alternative also the interaction between a height, and a (10–20%). The stand validation (Table 5) produced best
density variable proved to be significant in the models. results with the IM80 data, with an overall accuracy and
The validation of the Dirichlet models showed consistent kappa coefficient of 79% and 0.58%, respectively. The produ-
RMSEs for spruce, pine, and deciduous across ALS, IM80, and cer’s accuracy of classifying spruce was largest for ALS and
IM60 (Table 4). All the models for S% , P% , and D% had IM60 (86%), whereas for pine IM80 yielded the greatest accu-
RMSEs at plot and stand level smaller than 21.4 and 16.68, racy (73%). In the stand validation data there was no stand
respectively. The model using the IM80 dataset produced the where the deciduous class was dominant. However, one
smallest RMSE. Generally, the relative RMSE values for spruce, stand was classified as deciduous in IM80.
pine, and deciduous were 34.4%, 40.2%, and 76.6%, respect-
ively (Table 4). The pairwise t-test with Bonferroni correction
Total volume
revealed that there were no significant differences (p > .5) in
the predicted proportions of spruce, pine, and deciduous For all the point cloud datasets the volume models included
between ALS, IM80, and IM60. As shown in Figure 1 the only two structural variables (Table 6). The variables selected
overall trend for all three point cloud datasets was to overesti- in the ALS volume model were a height (p95) and a density
mate small proportions and underestimate large proportions. (d51) variable, whereas only height variables were included
As shown by a smaller slope in the regression lines (Figure 1), (p10 and p80) in the photogrammetric models. The adjusted
such effect was more pronounced for deciduous proportions, R 2 for the ALS, IM80, and IM60 datasets were 0.86, 0.83, and
and was of a similar magnitude for spruce and pine. 0.82, respectively, suggesting a fairly good model fit (Table 6).

Figure 1. Predicted versus ground reference stand level species proportions. The lines represent the 1:1 line (thicker line) and the different regression lines high-
lighting the species-wise differences in the prediction of the species proportions. The results are presented for each of the studied point cloud datasets: airborne laser
scanning (ALS), UltraCam XP imageries with 80% forward overlap (IM80) and 60% forward overlap (IM60). 173 × 62 mm (300 × 300 DPI).
SCANDINAVIAN JOURNAL OF FOREST RESEARCH 75

Table 5. Error matrices for the stand validation dataset. The error matrices for the plot leave one out cross validation are presented in parentheses. The results are
presented for each of the studied point cloud dataset: airborne laser scanning (ALS), UltraCam XP imageries with 80% forward overlap (IM80) and 60% forward
overlap (IM60).
Spruce Pine Deciduous Sum User’s accuracy
ALS
Spruce 32 (69) 11 (15) 0 (7) 43 (91) 0.74 (0.76)
Pine 5 (9) 15 (44) 0 (2) 20 (55) 0.75 (0.80)
Deciduous 0 (3) 0 (1) 0 (1) 0 (5) 0 (0.20)
Sum 37 (81) 26 (60) 0 (10) 63 (151)
Producer’s accuracy 0.86 (0.84) 0.57 (0.73) 0 (0.13) 0.74 (0.75)
Kappa 0.46 (0.54)
IM80
Spruce 31 (71) 7 (12) 0 (7) 38 (90) 0.81 (0.79)
Pine 5 (7) 19 (47) 0 (1) 24 (55) 0.78 (0.85)
Deciduous 1 (3) 0 (1) 0 (2) 1 (6) 0 (0.33)
Sum 37 (81) 26 (60) 0 (10) 63 (151)
Producer’s accuracy 0.84 (0.88) 0.73 (0.77) 0 (0.20) 0.79 (0.79)
Kappa 0.58 (0.61)
IM60
Spruce 32 (68) 10 (15) 0 (8) 42 (91) 0.76 (0.75)
Pine 5 (10) 16 (45) 0 (1) 21 (56) 0.75 (0.80)
Deciduous 0 (3) 0 (0) 0 (1) 0 (4) 0 (0.25)
Sum 37 (81) 26 (60) 0 (10) 63 (151)
Producer’s accuracy 0.86 (0.84) 0.61 (0.75) 0 (0.10) 0.75 (0.75)
Kappa 0.49 (0.54)

Table 6. Selected variables, estimated parameters, significance values, and Table 7. Mean difference (MD) and RMSE between estimated and ground
adjusted R 2 for the regression models for total volume (Vtot). The results are reference volume. Value in percentage of the ground reference value is
presented for each of the studied point cloud datasets: airborne laser presented in parentheses. The results are presented for each of the studied
scanning (ALS), UltraCam XP imageries with 80% forward overlap (IM80) and point cloud dataset: airborne laser scanning (ALS), UltraCam XP imageries
60% forward overlap (IM60). with 80% forward overlap (IM80) and 60% forward overlap (IM60).
Dependent Independent variablesa Adj. Dataset Validationa RMSE (m3 ha−1) MD (m3 ha−1)a
Dataset variable and intercept Estimate R2
ALS Plot 48.9 (19.7) 0.00 NS (0.00)
ALS log(Vtot) intercept 2.93760*** 0.86 Stand 31.0 (13.3) 9.86* (4.22)
p95 0.09079*** IM80 Plot 49.9 (20.1) 0.00 NS (0.00)
d51 1.49154*** Stand 31.2 (13.4) –0.9 NS (–0.4)
IM80 log(Vtot) intercept 3.96162*** 0.83 IM60 Plot 52.9 (21.3) 0.00 NS (0.00)
p10 0.06110*** Stand 35.9 (15.4) 1.5 NS (0.6)
p80 0.05882*** a
Level of significance: NS = not significant (>.05). *<.05. **<.01. ***<.001.
IM60 log(Vtot) intercept 3.93447*** 0.82 b
Plot level leave one out cross validation or stand level independent validation
p10 0.05834***
dataset.
p80 0.05655***
a
p10, p80, p95: percentiles of points’ height for 10, 80, and 95% (m); d51: canopy
density corresponding to the proportions of points above fraction no. 5. spruce, pine, and deciduous volumes were predicted at
*** p < .001. stand level with RMSE ranging from 46.6 (36.6%) to 51.9 m3
ha−1 (40.8%), 36.8 (46.5%) to 39.3m3 ha−1 (49.6%), and 23.0
(84.9%) to 24.0 m3 ha−1 (87.7%), respectively (Table 8). The
The RMSE was always smaller or equal to 52.9 (21.3%) and maximum relative MD was 3.5% (p > .05) for spruce, –10.8%
35.9 m3 ha−1 (15.4%) for the CV and stand level validation, for pine (p > .05), and 42.2% for the deciduous class (p
respectively. The maximum differences of relative RMSE < .01). Overall, the IM80 produced the most accurate results,
between the IM60 and the ALS models at plot and stand whereas the IM60 yielded the lowest accuracies. The decid-
level were 1.6 and 2.1 percentage points, respectively uous volume predictions were always overestimated by
(Table 7). According to the stand validation the ALS dataset 27.5–42.2%.
produced the smallest relative RMSE (13.3%), however, the
IM80 volume model resulted in the smallest relative MD
between ground reference and predicted values (–0.4%; Discussion
p > .05). The results of a two-sided t-test showed that for the
three volume models (ALS, IM80, and IM60) the stand predic- The results of the current study revealed that when applying
tions did not differ significantly (p > .51) between each other. the described method at stand level and by using highly over-
Furthermore, all models underestimated the volume when it lapping imagery (IM80), the species proportions were pre-
was larger than approximately 450 m3 ha−1 (Figure 2). dicted with an RMSE of maximum 16.8%, the stand
dominant species was classified with a maximum of 79%
accuracy. We also found that the total volume was predicted
with a relative RMSE of 13.4%, and species-specific volumes
Species-specific volume
were predicted with relative RMSE of 36.6%, 46.5%, and
At the plot level the VSS (Table 8) showed the lowest relative 84.9%, for spruce, pine, and deciduous, respectively. At
RMSE for spruce (43.8–50.3%), followed by pine (63.3– stand level the MDs between predictions obtained with the
68.8%), and deciduous species (90.9–113.8%). Similarly, photogrammetric models and ground reference values
76 S. PULITI ET AL.

Figure 2. Predicted and ground reference stand level total volume (m3 ha–1). The results are presented for each of the studied point cloud datasets: airborne laser
scanning (ALS), UltraCam XP imageries with 80% forward overlap (IM80) and 60% forward overlap (IM60). 172 × 61 mm (300 × 300 DPI).

Table 8. Mean difference (MD) and RMSE between estimated and ground method. One fundamental step in the present study was
reference species-specific volume. RMSE in percentage of the ground the prediction of species proportions using the spectral
reference value is presented in parentheses. The results are presented for
each of the studied point cloud dataset: airborne laser scanning (ALS), data. In this regard, one of the limitations of this study was
UltraCam XP imageries with 80% forward overlap (IM80) and 60% forward that it did not specifically accounted for variations in the spec-
overlap (IM60). tral data due to varying light and atmospheric conditions, het-
RMSE erogeneity of within-species reflectance properties, and view
Species Dataset Validationb (m3 ha−1) MD (m3 ha−1)a
angle geometry. Variations in light and atmospheric con-
Spruce ALS Plot 64.1 (43.8) –4.6 NS (–3.1)
Stand 47.9 (39.2) 4.5 NS (3.5) ditions were neglectable in the present study as the
IM80 Plot 71.5 (48.9) –2.3 NS (–1.6) imagery was collected with a single flight of approximately
Stand 46.6 (36.6) 0.2 NS (0.2) seven minutes. The within-species spectral variations and
IM60 Plot 73.7 (50.3) 4.7 NS (3.2)
Stand 51.9 (40.8) –3.88 NS (–3.0) the differences in spectral response due to view angle geome-
Pine ALS Plot 47.0 (66.4) –5.7 NS (–8.1) try were not studied as it was out of the scope of the present
Stand 37.4 (47.3) –5.4 NS (–6.8) research. In fact, up to date, regardless the research efforts to
IM80 Plot 44.8 (63.3) –4.5 NS (–6.4)
Stand 36.8 (46.5) –8.4 NS (–10.8) account for such sources of variability, there has not yet been
IM60 Plot 48.7 (68.8) –4.0 NS (–5.6) a satisfactory methodology for operational application.
Stand 39.3 (49.6) –6.0 NS (–7.5) Another possible limitation of the study was that the IM80
Deciduous ALS Plot 30.9 (100.1) 10.3*** (33.4)
Stand 23.7 (87.7) 10.1*** (39.7) point cloud was thinned by half and therefore the accuracy
IM80 Plot 35.2 (113.8) 6.8* (22.2) of the IM80 models possibly underestimated. However, it is
Stand 23.0 (84.9) 7.4** (27.5) also likely that increased point densities would not corre-
IM60 Plot 28.1 (90.9) 8.7*** (28.5)
Stand 24.0 (88.7) 11.3*** (42.0) spond to significant increase in accuracy. In fact, the increased
a
Level of significance: NS = not significant (>.05). *<.05. **<.01. ***<.001. point density would not affect the structural variables since
b
Plot level leave one out cross validation (CV) or stand level independent vali- the points’ vertical distribution was maintained when thin-
dation dataset (independent). ning. It is also likely that the spectral variables would not be
significantly affected by an increase in point density as the
average point spacing would be much larger than the
indicated that other than for the prediction of deciduous image ground sampling distance resulting in redundancy of
volume there were no significant systematic differences. the information when extracting the spectral information at
Unfortunately, no independent data were available for a point cloud level. The findings of the current study are consist-
thorough validation at plot level. Therefore, the plot level ent with previous research conducted using a combination of
results served here mostly for comparison purposes with pre- ALS and multispectral data both in terms of accuracy and in
vious studies. The increase in the MD between plot and stand regard to a decrease in accuracy for the less represented
level estimates was possibly attributable to an underestima- species in the studied area. The predicted species proportions
tion at plot level due to dependency issues in the cross vali- had a maximum RMSE of 21.4%, which is in line with what was
dation. Our findings also indicated that there was no shown by Ørka et al. (2013), who reported a maximum RMSE
significant difference in the results obtained with the three of 25% also at the plot level. A limitation of the Dirichlet
different point cloud datasets (ALS, IM80, and IM60). Since regression was the overestimation of low proportions and
the current study was an exploratory analysis targeted at eval- an underestimation of large proportions. This might be due
uating the performance of the method in relation to volume, a to the Dirichlet regression function, similarly to the beta
further evaluation on a wider variety of commonly used forest regression, does not allow the input of proportions equal to
biophysical properties for operational FMIs (e.g. Lorey’s 0 or 1 as dependent variables in the training dataset.
height, dominant height, basal area, and stem number) However, the sample plots included pure spruce plots (i.e.
should be carried out to explore the practical utility of the proportion = 1) or plots without any pine or deciduous (i.e.
SCANDINAVIAN JOURNAL OF FOREST RESEARCH 77

proportion = 0). The effect of the transformation (Equation (1)) coniferous-dominated boreal forest. Deciduous trees are
was a reduction in the proportion range in the training data often, growing sparsely and under partial cover. Such charac-
which was then reflected in the predictions by predicting teristics make it difficult to detect such trees especially from
only in the interval (0–1). Consistently with previous species- aerial imagery. We found that the prediction accuracy was
specific studies the accuracy decreased for the species with consistent with what was found by Maltamo et al. (2014b)
smaller proportions. The overall level of accuracy obtained using the K-MSN method. In fact, their results showed relative
using IM80 (79%) was similar to the ones reported by RMSE values of 17.3%, 29.6%, 54.8%, and 122.1% for total
Maltamo et al. (2014b) for manual classification accuracy volume (Vtot), spruce, pine, and deciduous volume, respect-
(82.5%), indicating that the method presented in this study ively. Compared to the results obtained when using a large
could represent an objective and cost-effective alternative number of sample plots (Packalén & Maltamo 2007), the
to traditional methods. Nevertheless, greater classification method explored in this paper yielded larger errors but fol-
accuracies were found by Maltamo et al. (2014b) and by lowed a similar trend of decreasing accuracy with decreasing
Ørka et al. (2013) when classifying dominant species using a species proportions. Packalén and Maltamo (2007) yielded a
random forest classifier, obtaining overall accuracies of species-specific volume relative RMSE of 28.1%, 32.6%, and
94.7% and 89.1%, respectively. Also, Packalén et al. (2009), 62.3% for pine, spruce, and deciduous, respectively, where,
using a larger number of sample plots and the K-MSN pine was the dominating species in terms of volume in the
method, yielded a greater accuracy in classifying dominant area, followed by spruce, and deciduous trees. Intrinsic differ-
species (92%). Generally, the dominant species classification ences in forest conditions, sampling design, number of
accuracy followed a similar pattern in all studies, with the sample plots, modeling, and validation approaches often
main species being the most accurately classified and the make the comparison of reliability figures between different
accuracy decreasing for the less represented species. The studies difficult (Maltamo et al. 2014b; Vauhkonen et al. 2014).
accuracy of the stand predicted volume ( Vtot ) was consistent A primary conclusion that can be drawn is that the use of
with the results of the study by Gobakken et al. (2014) who photogrammetric data in combination with the presented
used the same dataset, but different software for the photo- method was able to allocate timber volume to the different
grammetric processing. There was no significant difference tree species with similar accuracy to operational approaches.
between V  tot predicted using ALS, IM80, or IM60 (p > .5). As The Dirichlet regression proved to be a viable modeling
expected, the ALS yielded the greatest adjusted R 2 (0.86), approach to simultaneously predict the occurrence of more
and smallest RMSE equal to 31.0 m3 ha−1 (13.3%). In this than two species in a given landscape, even when the
regard, the main differences between the ALS and photo- number of field plots was small and therefore not optimal
grammetry-based Vtot models were that the first one used for non-parametric imputations. Regardless of the large
both a height and a density variable whereas the latter ones errors related to the prediction of the deciduous proportions,
used only height variables. The poor canopy penetration of it was encouraging to see that it was possible to derive
photogrammetric point clouds compared to ALS ones might species information of spruce and pine with a reasonable
explain the reason behind the lack of inclusion of density accuracy. The coniferous species held 88% of the volume in
metrics in the IM80 and IM60 volume models. However, it is the studied area, and represented approximately 98.5% of
important to note that in a previous study (Puliti et al. 2015) the commercial roundwood removals in Norway (Statistics
conducted in the same area using unmanned aerial system Norway 2016). The large errors related to the prediction of
photogrammetric data derived from imagery with large over- deciduous proportions therefore have a limited impact on
laps, one density metric was included in the volume model. the overall effectiveness of species-specific FMIs in Norwegian
This suggests that to an increase in the image overlap corre- boreal forests conditions.
sponds an increase in the capacity of photogrammetric data Photogrammetric data with 80% and 60% forward overlap
to acquire information closer to the ground in the presence were able to predict species-specific proportions and total
of canopy gaps. Therefore, we conclude that larger overlaps volume with similar accuracy as ALS, leading to a twofold con-
could potentially improve the performance photogrammetric clusion. First, we can affirm that the use of photogrammetric
models by providing a more heterogeneous vertical distri- data in modeling species-specific volume can be a valid
bution of the point cloud. Regardless of the importance of alternative to ALS. Second, as demonstrated in previous
species information in the prediction of volume, no predicted research with regard to total volume predictions, when
proportion (S% , P 
% , and D% ) was found to be significant in decreasing the forward overlap from 80% to 60% the
describing total volume. The poor correlation between the species proportions and total volume results were not signifi-
predicted species proportions and the volume may be cantly different, indicating the possibility to conduct this type
caused by a poor accuracy in plot level predictions of the of forest inventory without applying any change to imagery
Dirichlet model. Indeed, initial tests using ground reference acquisition standards, thus not increasing costs.
species proportions instead of predicted ones as the indepen- Future developments of the presented method should
dent variables for the Vtot models revealed that the deciduous focus on retrieving more representative spectral information
proportion was included for the ALS and IM80 models. There- from photogrammetric data. In fact, potential improvements
fore, more accurate predictions of the deciduous proportions in tree species discrimination can be achieved by retrieving
could potentially improve the models. The poor accuracy in more accurate point cloud color information, taking into
the predictions of deciduous proportions was possibly account for example directional reflectance (Korpela et al.
related to the small occurrence of this class in the 2014). The method should also be tested on a larger forest
78 S. PULITI ET AL.

area when the remotely sensed data are acquired under a stand properties derived from airborne laser scanner data. Can J For
variety of light and atmospheric conditions. These differences Res. 38:1095–1109.
Haara A, Haarala M. 2002. Tree species classification using semi-automatic
in the imagery are expected to affect both the quality of the
delineation of trees on aerial images. Scand J For Res. 17:556–565.
photogrammetric processing as well as the accuracy of the Hansen EH, Gobakken T, Bollandsås OM, Zahabu E, Næsset E. 2015.
Dirichlet regression due to a greater noise in spectral Modeling aboveground biomass in dense tropical submontane rainfor-
reflectance. est using airborne laser scanner data. Remote Sens. 7:788–807.
Haralick RM, Shanmugam K, Dinstein IH. 1973. Textural features for image
classification. IEEE Trans Syst Man Cybern SMC-3. 6:610–621.
Acknowledgements Henningsen A, Toomet O. 2011. Maxlik: a package for maximum likelihood
estimation in R. Comput Stat. 26:443–458.
Thanks to Blom Geomatics As for collection of the remote sensing data Hijazi RH, Jernigan RW. 2007. Modelling compositional data using Dirichlet
and processing of the airborne laser scanner data. The authors are also regression models. J Appl Probab Stat. 4:77–91.
grateful to Dr Ole Martin Bollandsås of NMBU for processing the field data. Holmgren J, Persson Å, Söderman U. 2008. Species identification of indi-
vidual trees by combining high resolution LiDAR data with multi-spec-
tral images. Int J Remote Sens. 29:1537–1552.
Disclosure statement Honkavaara E, Arbiol R, Markelin L, Martinez L, Cramer M, Bovet S,
Chandelier L, Ilves R, Klonus S, Marshal P, et al. 2009. Digital airborne
No potential conflict of interest was reported by the authors. photogrammetry—a new tool for quantitative remote sensing? A
state-of-the-art review on radiometric aspects of digital photogram-
metric images. Remote Sens. 1:577.
References Järnstedt J, Pekkarinen A, Tuominen S, Ginzler C, Holopainen M, Viitala R.
2012. Forest variable estimation using a high-resolution digital surface
Agisoft LLC. 2014. Agisoft photoscan user manual. St. Petersburg. model. ISPRS J Photogramm Remote Sens. 74:78–84.
Available from: http://www.agisoft.com/pdf/photoscan-pro_1_1_en. Korhonen L, Korhonen KT, Stenberg P, Maltamo M, Rautiainen M. 2007.
pdf Local models for forest canopy cover with beta regression. Silva
Bohlin J, Olsson H, Olofsson K, Wallerman J. 2006. Tree species discrimi- Fenn. 41:671–685.
nation by aid of template matching applied to digital air photos. In: Korpela I, Heikkinen V, Honkavaara E, Rohrbach F, Tokola T. 2011. Variation
International workshop on 3D remote sensing in forestry; Vienna: and directional anisotropy of reflectance at the crown scale—impli-
BOKU. Available from http://www.rali.boku.ac.at/fileadmin/data/ cations for tree species classification in digital aerial images. Remote
H03000/H85000/H85700/workshops/3drsforestry/presentations/7.4- Sens Environ. 115:2062–2074.
olsson.pdf Korpela I, Mehtätalo L, Markelin L, Seppänen A, Kangas A. 2014. Tree
Bohlin J, Wallerman J, Fransson JES. 2012a. Forest variable estimation species identification in aerial image data using directional reflectance
using photogrammetric matching of digital aerial images in combi- signatures. Silva Fenn. 48:1087.
nation with a high-resolution DEM. Scand J For Res. 27:692–699. Leckie D, Gougeon F, Hill D, Quinn R, Armstrong L, Shreenan R. 2003.
Bohlin J, Wallerman J, Fransson JES. 2012b. Species-specific forest variable Combined high-density LiDAR and multispectral imagery for individual
estimation using non-parametric modelling of multi-spectral photo- tree crown analysis. Can J Remote Sens. 29:633–649.
grammetric point cloud data. In: International archives of the photo- Lumley T, Miller A. 2009. Leaps: regression subset selection. Available
grammetry, XXII ISPRS Congress; Melbourne. Available from: http:// from: http://cran.r-project.org/web/packages/leaps/leaps.pdf
pub.epsilon.slu.se/10635/8/bohlin_et_al_130710.pdf Maier MJ. 2014. Dirichletreg : Dirichlet regression for compositional data in
Braastad H. 1966. Volume tables for birch. Meddr norske SkogforsVes. R Dirichletreg: Dirichlet regression for compositional data in R. Version
21:23–78. Norwegian with English summary. 0.4.0; Vienna: WU Vienna University of Economics and Business,
Brantseg A. 1967. Volume functions and tables for scots pine. South Department of Statistics and Mathematics.
Norway. Meddr norske SkogforsVes. 22:689–739. Norwegian with Maltamo M, Næsset E, Vauhkonen J. 2014a. Forestry applications of air-
English summary. borne laser scanning. Springer.
Dalponte M, Ørka HO, Ene LT, Gobakken T, Næsset E. 2014. Tree crown Maltamo M, Ørka HO, Bollandsås OM, Gobakken T, Næsset E. 2014b. Using
delineation and tree species classification in boreal forests using pre-classification to improve the accuracy of species-specific forest
Hyperspectral and ALS Data. Remote Sens Environ. 140:306–317. attribute estimates from airborne laser scanner data and aerial
Dalponte M, Orka HO, Gobakken T, Gianelle D, Næsset E. 2013. Tree images. Scand J For Res. 30:1–22.
species classification in boreal forests with hyperspectral data. IEEE Maltamo M, Packalén P, Kangas J, Uuttera J, Heikkila J. 2011. Airborne
Trans Geosci Remote Sens. 51:2632–2645. laser scanning based stand level management inventory in Finland.
Donoghue D, Watt P, Cox N, Wilson J . 2007. Remote sensing of species In: Proceedings of SilviLaser 2011, 11th International Conference
mixtures in conifer plantations using LiDAR height and intensity on LiDAR Applications for Assessing (pp. 16–19). University of
data. Remote Sens Environ. 110:509–522. Tasmania.
Eid T, Næsset E. 1998. Determination of stand volume in practical forest Mikhail EM, Bethel JS, McGlone JC. 2001. Introduction to modern photo-
inventories based on field measurements and photo-interpretation: grammetry. New York, NY: Wiley.
the Norwegian experience. Scand J For Res. 13:246–254. Miller RG. 1981. Simultaneous statistical inference. (2nd ed.). New York, NY:
Erikson M. 2004. Species classification of individually segmented tree Springer-Verlag.
crowns in high-resolution aerial images using radiometric and morpho- Mora B, Wulder MA, White JC. 2010. Identifying leading species using tree
logic image measures. Remote Sens Environ. 91:469–477. crown metrics derived from very high spatial resolution imagery in a
Fitje A, Vestjordet E. 1977. Stand height curves and new tariff tables for boreal forest environment. Can J Remote Sens. 36:332–344.
Norway spruce. Meddr norske SkogforsVes. 34:23–62. Norwegian Nilsson M. 1996. Estimation of tree heights and stand volume using an air-
with English summary. borne lidar system. Remote Sens Environ. 56:1–7.
Franklin SE, Hall RJ, Moskal LM, Maudie AJ, Lavigne MB. 2000. Niska H, Skön J-P, Packalén P, Tokola T, Maltamo M, Kolehmainen M. 2010.
Incorporating texture into classification of forest species composition Neural networks for the prediction of species-specific plot volumes
from airborne multispectral images. Int J Remote Sens. 21:61–79. using airborne laser scanning and aerial photographs. IEEE Trans
Gobakken T, Bollandsås OM, Næsset E. 2014. Comparing biophysical forest Geosci Remote Sens. 48:1076–1085.
characteristics estimated from photogrammetric matching of aerial Nurminen K, Karjalainen M, Yu X, Hyyppä J, Honkavaara E. 2013.
images and airborne laser scanning data. Scand J For Res. 30:1–37. Performance of dense digital surface models based on image matching
Gobakken T, Næsset E. 2008. Assessing effects of laser point density, in the estimation of plot-level forest variables. ISPRS J Photogramm
ground sampling intensity, and field sample plot size on biophysical Remote Sens. 83:104–115.
SCANDINAVIAN JOURNAL OF FOREST RESEARCH 79

Næsset E. 2002. Predicting forest stand characteristics with airborne scan- Straub C, Stepper C, Seitz R, Waser LT. 2013. Potential of ultracamx stereo
ning laser using a practical two-stage procedure and field data. Remote images for estimating timber volume and basal area at the plot level in
Sens Environ. 80:88–99. mixed European forests. Can J For Res. 43:731–741.
Næsset E. 2004a. Accuracy of forest inventory using airborne laser scan- Terrasolid. 2012. Terrascan user’s guide, Terrasolid Ltd., Helsinki. Available
ning: evaluating the first Nordic full-scale operational project. Scand J from: https://www.terrasolid.com/download/tscan.pdf
For Res. 19:554–557. Vastaranta M, Wulder MA, White JC, Pekkarinen A, Tuominen S, Ginzler C,
Næsset E. 2004b. Practical large-scale forest stand inventory using a small-- Kankare V, Holopainen M, Hyyppa J, Hyyppa H. 2013. Airborne laser
footprint airborne scanning laser. Scand J For Res. 19:164–179. scanning and digital stereo imagery measures of forest structure: com-
Næsset E. 2005. Assessing sensor effects and effects of leaf-off and leaf-on parative results and implications to forest mapping and inventory
canopy conditions on biophysical stand properties derived from small- update. Can J Remote Sens. 39:382–395.
footprint airborne laser data. Remote Sens Environ. 98:356–370. Vauhkonen J, Ørka HO, Holmgren J, Dalponte M, Koch B. 2014. Tree
Næsset E. 2014. Area-based inventory in Norway – from innovation to species recognition based on airborne laser scanning and complemen-
operational reality. In: Maltamo M, Næsset E, Vauhkonen J, editors. tary data sources. In: Maltamo M, Næsset E, Vauhkonen J, editors.
Forestry applications of airborne laser scanning. Dordrecht: Springer. Forestry applications of airborne laser scanning – concepts and case
p. 215–240. studies. Dordrecht: Springer. p. 135–156.
Ørka HO, Dalponte M, Gobakken T, Næsset E, Ene LT. 2013. Characterizing Vestjordet E. 1967. Functions and tables for volume of standing trees.
forest species composition using multiple remote sensing data sources Norway Spruce. Meddr norske SkogforsVes. 22:539–574. Norwegian
and inventory approaches. Scand J For Res. 28:677–688. with English summary.
Ørka HO, Gobakken T, Næsset E, Ene L, Lien V. 2012. Simultaneously Vestjordet E. 1968. Volum Av Nyttbart Virke Hos Gran Og Furu Basert På
acquired airborne laser scanning and multispectral imagery for individ- Relativ Høyde Og Diameter I Brysthøyde Eller Ved 2,5 M Fra
ual tree species identification. Can J Remote Sens. 38:125–138. Stubbeavskjær [Merchantable volume of Norway spruce and scots
Ørka HO, Næsset E, Bollandsås OM. 2010. Effects of different sensors and pine based on relative height and diameter at breast height or 2.5 M
leaf-on and leaf-off canopy conditions on echo distributions and indi- above stump level]. Meddr norske SkogforsVes. 25:411–459.
vidual tree properties derived from airborne laser scanning. Remote Norwegian with English summary.
Sens Environ. 114:1445–1461. Vihervaara P, Mononen L, Auvinen A-P, Virkkala R, Lü Y, Pippuri I, Packalen
Packalén P, Maltamo M. 2007. The k-msn method for the prediction of P, Valbuena R. 2015. How to integrate remotely sensed data and biodi-
species-specific stand attributes using airborne laser scanning and versity for ecosystem assessments at landscape scale. Landsc Ecol.
aerial photographs. Remote Sens Environ. 109:328–341. 30:501–516.
Packalén P, Suvanto A, Maltamo M. 2009. A two stage method to estimate Warner TA, McGraw JB, Landenberger R. 2006. Segmentation and classifi-
species-specific growing stock. Photogramm Eng Remote Sens. cation of high resolution imagery for mapping individual species in a
75:1451–1460. closed canopy, deciduous forest. Sci China Ser E. 49:128–139.
Puliti S, Ørka HO, Gobakken T, Næsset E. 2015. Inventory of small forest Waser LT, Ginzler C, Kuechler M, Baltsavias E, Hurni L. 2011. Semi-auto-
areas using an unmanned aerial system. Remote Sens. 7:9632. matic classification of tree species in different forest ecosystems by
Rahlf J, Breidenbach J, Solberg S, Næsset E, Astrup R. 2014. Comparison of spectral and geometric variables derived from airborne digital sensor
four types of 3d data for timber volume estimation. Remote Sens (Ads40) and Rc30 data. Remote Sens Environ. 115:76–85.
Environ. 155:325–333. Wing BM, Ritchie MW, Boston K, Cohen WB, Gitelman A, Olsen MJ.
R Core Team. 2015. A language and environment for statistical computing. 2012. Prediction of understory vegetation cover with airborne
R Foundation for Statistical Computing. Vienna. Available from: http:// lidar in an interior ponderosa pine forest. Remote Sens Environ.
www.R-project.org 124:730–741.
Smithson M, Verkuilen J. 2006. A better lemon squeezer? Maximum-like- Wu C. 2004. Normalized spectral mixture analysis for monitoring
lihood regression with beta-distributed dependent variables. Psychol urban composition using etm+imagery. Remote Sens Environ.
Methods. 11:54–71. 93:480–492.
Snowdon P. 1991. A ratio estimator for bias correction in logarithmic Yu B, Member S, Ostland IM, Gong P, Pu R. 1999. Penalized discriminant
regressions. Can J For Res. 21:720–724. analysis of in situ hyperspectral data for conifer species recognition.
Statistics Norway. 2016. Commercial roundwood removals, Q1 2016, pre- IEEE Trans Geosci Remote Sens. 35:2569–2577.
liminary figures; cited [2016 Apr 20 ]. Available from: http://www.ssb. Zvoleff A. 2015. Package ‘GLCM’. version 1.2. Available from: http://cran.r-
no/en/jord-skog-jakt-og-fiskeri/statistikker/skogav/kvartal project.org/web/packages/glcm/glcm.pdf

You might also like