2003). Some
PS
technologies have been optimized for forestry,
such as ForeStereo (a camera system with fish-eye lenses
(Montes
et al.
2009)) and Terrestrial Laser Scanning (
TLS
)
(Liang
et al.
2016; Newnham
et al.
2015).
ForeStereo is a passive optical sensor composed of two
hemispherical cameras capturing the visible radiation re-
flected by the surface of surrounding objects (e.g. trees) in
a single shot. The stereoscopic hemispherical images were
firstly used for measurement of tree diameter, height, and
volume by Rodríguez-Garcia
et al.
(2014). ForeStereo technol-
ogy has been incorporated into the Spanish National Forest
Inventory (Sánchez González
et al.
2016) and National Parks
forest monitoring in Spain (Sangüesa-Barreda
et al.
2015;
Rubio-Cuadrado
et al.
2018a; Rubio-Cuadrado
et al.
2018b).
ForeStereo is widely used as part of the field survey in natural
forests and plantations of different species in Spain and it is
being tested in Brazil and Chile. Computer vision techniques
have been developed for the segmentation of ForeStereo
images to identify trees (Sánchez-González
et al.
2016) and
a methodology has been proposed for the three-dimensional
(3D) reconstruction of stem profiles from the angular dispar-
ity between the matched stem sections in both stereoscopic
images and epipolarity restrictions (Rodríguez-García
et al.
2014). ForeStereo has been specifically developed for for-
estry and employs software designed specifically to estimate
stand variables such as basal area (
BA
) or number of trees per
hectare (
N
) (Sánchez-González
et al.
2016). Applications for
vertical distribution of forest fuel assessment (Marino
et al.
2018) or species classification (Gea-Izquierdo
et al.
2015) have
also been developed.
Terrestrial Laser Scanner used in forest inventories
provides a dense point cloud that enables the 3D structural
reconstruction (Strahler
et al.
2008; Yao
et al.
2011) and
estimation of forest inventory variables such as
N
,
BA
, volume
or biomass (Lovell
et al.
2011; Astrup
et al.
2014; Liang
et
al.
2016).
TLS
provides detailed data of below-canopy forest
structure that can be used for precise biomass estimations and
development of allometric equations, as well as crown shape
and tree morphology modelling (Côte
et al.
20
et al.
2013; Hauglin
et al.
2013). However, the
beam returns generated by
TLS
requires compl
cessing to select returns from stems, foliage or
fit geometrical models for diameter and volume estimation
(Newnham
et al.
2015). In addition,
TLS
devices are expensive
and its use for the field stage in large scale forest inventories
usually is not justified (White
et al.
2016). Modern
TLS
de-
vices reduce weight, and two-dimensional (2D) laser scanner
offers an easy to handle solution, but data is limited to the
sensors plane (Ringdahl
et al.
2013; Brunner and Gizachew
2014). Compared with
TLS
, photogrammetric
PS
techniques
(including ForeStereo) have better battery performance and
reduced device weight, and faster image processing through
computer vision techniques (Herrera
et al.
2009; Herrera
et al.
2011). Photogrammetric
PS
technologies employing conven-
tional lenses require multiple image acquisitions (Clark
et al.
2000; Varjo
et al.
2006; Forsman
et al.
2012), and processing
becomes complicated. ForeStereo employs fisheye lenses with
180
o
field of view, so all features above the horizontal plane
passing through the lens are projected in the image, enabling
the 3D forest structural reconstruction around the sampling
point from a single pair of images.
Large-scale operational forest inventories require a large
number of sampling points covering the target area, therefore
single scan plots are an economically advantageous option
(Ducey
et al.
2014). In order to become functional, a
PS
instru-
ment for use in operational forest inventories should pro-
vide repeatable measures of the forest variables in a fast and
economic manner, as well as a measure of the accuracy and
reliability of the acquired data. For estimation of stand level
variables like
BA
or
N
, the fixed-point techniques are based on
the angle count sampling method, which is the basis of the
Relaskop (Bitterlich 1947). In the angle count sampling meth-
od, those stems projected in the visor which are wider than
a reference threshold are sampled, so the maximum distance
of detection depends on the size of each tree. Fixed point
sampling typically underestimates the number of trees for two
main reasons: (1) the instrument bias derived from a limited
range of detection, which depends on the resolution of the
scan or image and the diameter of trees (Seidel and Ammer
2014) and (2) the process of tree retrieval from a single scan
or image: some trees actually present in the plot are occluded
by others, hence reducing the tree detection rate (Lovell
et al.
2011). The underestimation issue becomes more important
in very dense forests and has been analyzed in recent studies
(Yang
et al.
2013; Brunner and Gizachew 2014). Three main
methods have been proposed to correct the effects of tree
occlusion on the estimation of plot and stand level variables,
leading to less biased estimations:
i) Attenuation model (Strahler
et al.
2008). This method
assumes that the gap probability—where a gap is the area
between trees that allows clear visibility to more distant
trees—decreases exponentially as a function of the dis-
tance to the sensor and an effective diameter of trees, fol-
lowing a Poisson model. The effective diameter includes
branches, foliage, shrubs, and any other intercepting ele-
ments. This method has been used for counting stems and
for estimation of biomass from
TLS
data (Yao
et al.
2011).
ii) Modelling a detection function based on distance-sam-
pling (Ducey
et al.
2013). In this case, the probability of
detecting a tree depends on its distance from the sensor.
The function to model probabilities is fitted to the dis-
tribution of distances of trees that are actually detected
(apparent trees). Distance-sampling methods are widely
used in ecology, where the mathematical framework for
population summary statistics and variance estimation
has been established (Buckland
et al.
2001). Astrup
et al.
d two functions to model the probability of
amely Half-Normal and Hazard-Rate, and
es with data from
TLS
.
e proportion of sampling area shadowed
by the apparent trees (Seidel and Ammer 2014). The plot
area shadowed by trees can be calculated from the diam-
eter and the actual distance from the sensor of all retrieved
trees. This method was adapted by Sánchez-González
et
al.
(2016) for ForeStereo.
These methods have been tested with
TLS
data mainly focused
on plantations of even-aged stands managed for timber pro-
duction, but comparative works using ForeStereo data are
lacking. The first two methods depend on parameters linked
to the stand conditions (density, beam interception by foliage
or shrubs, and other features hampering the detection of trees)
and may show limitations in natural heterogeneous forests,
where the method proposed by Seidel and Ammer (2014)
seems more suitable. Under the last method the apparent
trees generate a shadowed angular sector behind them. The
shadowed area is then subtracted from the sampling plot area.
However, this method may underestimate the number of trees
of smaller Diameter at Breast Height (
DBH
) classes for which
the probability of nondetection is greater (Sánchez-Gonzalez
et al.
2016). Moreover, the method proposed by Seidel and
Ammer (2014) must be adapted for use with ForeStereo
because in the case of stereoscopic hemispherical images,
mensuration requires correspondence between the projection
of the tree in both images of the stereoscopic pair, so occlu-
sion in either of the two images hampers tree detection.
494
July 2019
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
