PE&RS August 2015 - page 627

A minimum pulse density of 10 pl m-2 was stipulated prior
to capture. By maintaining swath overlap in postprocessing
acquisition pulse density is approximately doubled, though
variation in density still occurred. Although capture over
a large area at this density with a 50 percent swath overlap
would not be considered operational for a land management
agency (Wulder et al., 2012b), oversampling allows a method-
ical simulation of various pulse densities.
For each study area, 50 plot centers were randomly located
across the 5 km × 5 km capture footprint. To ensure spatial
independence of extracted plots, plot centers were located at
a minimum distance of 250 m apart. From each plot centre,
ALS data for a circular plot with a 25 m radius was extracted.
To decrease the risk of duplication of returns in a thinned da-
taset, extracted plots with a pulse density of <10 pl m-2 were
rejected and new plots selected. Pulse density of plots was
estimated by totaling the number of first returns (e.g., points
with a return number metadata value of 1) per plot and divid-
ing by plot area (Jakubowski et al., 2013).
Data Processing
There are a number of existing studies that assess the ac-
curacy of forest metric retrieval from point clouds of differ-
ent densities. Ideally, analysis of different pulse densities
would be undertaken on datasets captured at different flying
altitudes, or aircraft speeds (Goodwin et al., 2006; Magnus-
son et al., 2007; Morsdorf et al., 2008; Takahashi et al., 2008;
Thomas et al., 2006). This permits analysis of additional
variables altered by changing capture specifications such
as instantaneous laser pulse power and laser footprint size.
However, capture at multiple altitudes and aircraft speeds
is limited by cost, particularly over large or discontinuous
study areas such as in this study, and alternative modeling
methods are required. A number of authors have decreased
the number of points in a dataset to match a required point
density (Maltamo et al., 2006; Tesfamichael et al., 2010; Watt
et al., 2013); however this does not necessarily replicate the
reduction in pulse density caused by change in altitude or
aircraft speed if simulating anything other than a first-return
dataset (Jakubowski et al., 2013). The majority of techniques
that simulate a reduction in pulse density do so by superim-
posing a regular grid over the study area of a specified spatial
resolution to attain required pulse densities, returns are then
randomly selected from within each voxel (Gobakken and
Næsset, 2008; Jakubowski et al., 2013; Korhonen et al., 2011;
Næsset, 2009). However, application of this technique may
not replicate the regular scan pattern in which data is col-
lected (Baltsavias, 1999), particularly when simulating low
pulse densities. Other techniques include the removal of
alternate pulses and scan lines (Treitz et al., 2012), stipulating
a minimum horizontal distance between returns (Magnusson
et al., 2007) or systematically thinning a dataset utilizing GPS
time (Khosravipour et al., 2014); albeit these techniques are
only suitable for generating single or first-return datasets.
Here we introduce a new technique that can (a) systemati-
cally sample the original dataset to allow analysis of sampling
variance and acquisition reproducibility, and (b) simulate
multi-return capture (e.g., as opposed to first or first-and-last
capture) that state-of-the-art small-footprint instruments can
produce. Computations were carried out with the ForestLAS
Python module (contact corresponding author) except where
stipulated. For the 250 extracted plots, nine different plotwise
realizations were simulated at six different pulse densities (Ta-
ble 3). This process resulted in a total of 15,500 simulations.
T able 1. Description and Location of Study S ites Sorted by M ean Canopy Height
Site Coordinates Elevation Mean slope Description
Canopy
height
Mean annual
rainfall^
Watts Creek (WC) 37° 41' 22"
145° 41' 5" 975 – 1220 m 15° Very tall, dense and highly productive forest.
Tall and mature mid- and under-storey. 20 – 90 m 1630 mm
Robsons Creek (RC) 17° 6' 20"
145° 37' 16"
1160 – 700 m 24° Notophyll vine forest with a tall canopy.
High species diversity. 25 – 40 m 1890 mm
Zig Zag Creek (ZZ) 37° 28' 26"
148° 20' 19" 200 – 400 m 16°
Medium height, dense forest characterised
by variability and tree species diversity.
Dense understorey.
15 – 45 m 820 mm
Litchfield (LI) 13° 10' 39"
130° 47' 23" 210 – 230 m 3° Savanna, eucalypt open forest. 10 – 25 m 1370 mm
Rushworth Forest (RF) 36° 45' 11"
144° 57' 57" 170 – 300 m 5° Short, open forest with low canopy density
and a sparse, shrubby understorey. 10 – 20 m 580 mm
Credo (CR) 30° 11' 22"
120° 39' 17"
440 – 480 m 2° Open woodland inter-dispersed with open,
treeless areas. Small shrub layer prevalent. 5 – 25 m 260 mm
^ Rainfall data from Hijmans et al. (2005)
Table 2.Flight and Sensor S pecifications for the ALS Acquisitions
Specifications
Capture specifications
Flying height 300 m agl
Target pulse density (excl. overlap) 10 pulses m-2
Absolute vertical accuracy <0.15 m
Absolute horizontal accuracy <0.15 m
Mean footprint diameter 0.15 m
Instrument specifications
Instrument Riegl LMS-Q560 laser scanner
(Horn, Austria)
Operating wavelength 1550 nm
Beam divergence 0.5 mrad
Max off-nadir scan angle ±22.5°
Outgoing pulse rate 240 kHz
T able 3. Pulse Density and Required Pulse Spacing
Pulse density (pulse m-2 ) Pulse spacing (m)
0.05 4.47
0.1 3.16
0.5 1.41
1 1.00
2 0.71
4 0.5
10 0.32
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING August 2015 627