Occlusion Probability in Operational Forest
Inventory Field Sampling with ForeStereo
F. Montes, A. Rubio-Cuadrado, M. Sánchez-González, I. Aulló-Maestro, M. Cabrera, and C. Gómez
Abstract
Field data in forest inventories are increasingly obtained
using proximal sensing technologies, often under fixed-point
sampling. Under fixed-point sampling some trees are not
detected due to instrument bias and occlusions, hence involv-
ing an underestimation of the number of trees per hectare
(N). The aim here is to evaluate various approaches to correct
tree occlusions and instrument bias estimates calculated
with data from ForeStereo (proximal sensor based on stereo-
scopic hemispherical images) under a fixed-point sampling
strategy. Distance-sampling and the new hemispherical
photogrammetric correction (
HPC
), which combines image
segmentation-based correction for instrument bias with a
novel approach for estimating the proportion of shadowed
sampling area in stereoscopic hemispherical images, best
estimated N and basal area (
BA
). Distance-sampling slightly
overestimated N (11% bias, 0.60 Pearson coefficient with the
reference measures) and
BA
(4%, 0.82).
HPC
provided less
biased N estimates (-6%, 0.61) but underestimated
BA
(-8%,
0.83).
HPC
most accurately retrieved the diameter distribution.
Introduction
Understanding forests dynamics is necessary for sustainable
forest management. Forest inventories supported by data
measured on the ground and data acquired remotely via aerial
or satellite platforms enable the monitoring of forest structure,
growth, change in composition, responses t
traits, and decline originated by climatic or
Continuous development of technology for
processing, and analysis contributes to imp
of forest inventories. Remote sensing provides information on
forest variables such as cover (Morsdorf
et al.
2006), struc-
ture (Gómez
et al.
2011), volume (Vauhkonen
et al.
2011), or
biomass (Næsset and Gobakken 2008), with complete spa-
tial coverage. A wide range of sensors with different spatial
resolutions (from hundreds to less than 1 m) are suitable for
forest applications (White
et al.
2016). Satellite imagery and
aerial Light Detection and Ranging (
LiDAR
) data support map-
ping and updating National Forest Inventory estimations (e.g.
in Finland, (Tomppo
et al.
2008) or in Canada (Hilker
et al.
2008)). Changes in land cover and forest variables like bio-
mass are mapped and monitored with optical and radar satel-
lite sensors (White
et al.
2017; Matasci
et al.
2018; Santoro
et al.
2018). For height retrieval and assessment of vertical
structure, Synthetic Aperture Radar may be employed by itself
(e.g. Garestier
et al.
2008; Tebaldini and Rocca 2012) and its
combination with
LiDAR
data is expected to provide improved
resolution and accuracy for large scale assessments (Qi
et al.
2019). For estimation of forest inventory variables at regional
scale
LiDAR
is the most precise technology (Wulder
et al.
2013)
enabling structural characterization (Valbuena
et al.
2013;
Bottalico
et al.
2017) and measurement of individual tree at-
tributes (Hauglin
et al.
2014).
LiDAR
provides precise digital
elevation models and metrics for characterization of height
distribution (Lindberg
et al.
2012) and is increasingly being
used to generalize the sampling plot data to the whole area
through area-based models (Næsset 2002; Bouvier
et al.
2015).
Research is ongoing to overcome
LiDAR
limitations in the esti-
mation of tree diameter distribution (Magnussen
et al.
2013)
or species composition (Maltamo
et al.
2009), which may
benefit from the combination with other sensors (Holmgren
et
al.
2008; Puttonen
et al.
2010; Zhao
et al.
2018). Forest canopy
surface height can also be estimated comparing image-based
point clouds derived from digital aerial photography and
terrain elevation models of high spatial resolution generated
with
LiDAR
(Leberl
et al.
2010). Digital aerial photography is
less costly than
LiDAR
(White
et al.
2013) and provide spectral
information to support species classification (St-Onge
et al.
2015); however, image-based point clouds do not penetrate
through tree crowns in dense stands (White
et al.
2013) be-
ing
LiDAR
more informative for forest inventory applications.
Equipped with optical sensors unmanned aerial vehicles
(
UAVs
) can retrieve photogrammetric surface models at local
high spatial resolution, and spectral vegeta-
upport species classification (Tuominen
et al.
practical concerns in the use of
UAVs
include
, power autonomy, payload weight, and local
regulations (Gómez and Green, 2017; Manfreda
et al.
2018).
Although the contribution of remotely sensed data to forest
inventories is constantly increasing, field data is needed to
calibrate and validate the models (Brosofske
et al.
2014). Field
data is expensive to acquire, therefore, a sampling survey ap-
proach is generally used (Mandallaz and Ye 1999). Proximal
Sensing (
PS
) techniques, which refer to the acquisition of data
with a sensor from a relatively short distance, may provide
information on variables which are complex to measure
manually, such as spatial arrangement and dimensions of
trees (Rodríguez-García
et al.
2014), canopy features (Seidel
et
al.
2012), stand volume (Astrup
et al.
2014), or forest fuel (Pi-
mont
et al.
2015), complementing the field measurement and
enhancing performance (Dassot
et al.
2011). The combination
of remote sensing and proximal sensing may improve estima-
tions and reduce the cost of inventories (Lindberg
et al.
2012).
The development of point sampling
PS
technologies—
based on laser distance or photogrammetry—to measure tree
dimensions and to estimate forest parameters at plot or stand
level began in the early 2000s (Clark
et al.
2000; Lovell
et al.
Fernando Montes, Mariola Sánchez-González, Isabel Aulló-
Maestro, and Cristina Gómez are with INIA-CIFOR, Ctra. de la
Coruña km 7.5, 28040 Madrid, Spain (
.
Álvaro Rubio-Cuadrado is with the Departamento de Sistemas y
Recursos Naturales, Escuela Técnica Superior de Ingeniería de
Montes, Forestal y del Medio Natural, Universidad Politécnica
de Madrid, Ciudad Universitaria, 28040, Madrid, Spain.
Miguel Cabrera is with Aranzada Gestión Forestal, C/Alonso
Heredia 31, 28028, Madrid, Spain.
Photogrammetric Engineering & Remote Sensing
Vol. 85, No. 7, July 2019, pp. 493–508.
0099-1112/19/493–508
© 2019 American Society for Photogrammetry
and Remote Sensing
doi: 10.14358/PERS.85.7.493
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
July 2019
493
