stratifications will yield different results for different sites
(provided their physical characteristics differ) whereas
GLAS
waveform data recorded at different locations are still subject
to all data stratification techniques investigated here (laser
number, transmission energy, and phenology)
.
Of the three study sites investigated, Watts Creek yields the
best overall result for all canopy height metrics tested, while
Tumbarumba performs better than Robson Creek, suggest-
ing that
GLAS
provides better height estimates for more open
forests than for dense, tropical rainforests. This is consistent
with existing literature (Simard
et al
., 2011; Los
et al
., 2012;
Mahoney
et al
., 2014), highlighting that heterogeneity, and
tall, dense canopies are problematic for
GLAS
, whereas the
relative homogeneity in vegetation structure present across
the Eucalypt sites is more accurately captured in
GLAS
data.
Optimal GLAS Dataset
Best comparison results are achieved where
GLAS
data are
acquired during summertime phenological conditions, from
a high energy (>28 mJ) laser 3 transmission; the most highly
correlated
GLAS
/
ALS
results are achieved from the 95
th
ALS
height percentile calculated directly from an all return point
cloud. A side-by-side statistical comparison of non-optimized
(p95 controls), and optimized
GLAS
/
ALS
comparisons are
noted in Table 8; comparisons are also illustrated in Figure 9
.
A 15 percent improvement is noted between control and
optimized data comparisons for
RMSE
, whereas a 26 percent
improvement is noted between values of R
2
. The smallest
improvement is noted between F
20
, at 6 percent, where a small
deviation from the best achievable F
B
(0.01) is also noted. Ad-
ditionally, where the canopy is less than 20 m, and greater than
40 m, greater divergence in confidence intervals is noted for
the optimized dataset when compared to the control dataset.
This is an artifact of the lack of data present in these regions
.
Results suggest that laser number selection yields the greatest
improvements with regards to
GLAS
/
ALS
comparisons of canopy
height. This result is based on the noted change in comparison
statistics from the statistically best control dataset and the statis-
tically best dataset within each test. By the same procedure, the
phenological state at the time of
GLAS
data acquisition appears
to affect canopy height estimates the least. The change in sta-
tistics between the best control dataset and the best datasets of
each test are noted in Table 9, expressed as percentage change.
Discussion
The sensitivity of canopy heights retrieved from
GLAS
data
have been investigated with respect to data stratifications in-
duced by system, temporal, and spatial sources; results were
assessed against spatially concurrent
ALS
data. Investigations
Figure 6. Stratified control comparisons for the best GLAS height method and ALS pXX as a function of ALS data source: (a) all returns, (b)
first returns, and (c) raster; and each phenological state: i) summer foliage (April to September), and ii) winter foliage (October to March).
Linear models and associated 95 percent confidence intervals are illustrated.
T
able
5. S
ummary
S
tatistics
for
C
ontrols when
F
iltered
A
ccording
to
T
ime
of
GLAS D
ata
A
cquisition
(S
ummer
/W
inter
P
henology
);
p
XX
is
the
M
ost
C
losely
R
elated
ALS H
eight
P
ercentile with
E
ither
GLAS H
eight
D
erivation
;
the
B
est
O
verall
ALS-GLAS R
elationship
is
H
ighlighted
for
E
ach
ALS D
ata
S
ource
.
N
ote
: S
is
the
phenological
S
tate
at
D
ata
A
cquisition
,
where
SP
is
S
ummertime
and
WP W
intertime
.
RH
100
RH
ROS
Data S N
RMSE
R
2
F
20
F
B
pXX ∇
RMSE
R
2
F
20
F
B
pXX
Raster
SP
103
1.05
12.66
0.35 0.68 0.05 p99
0.91
11.01
0.45 0.68 0.09 p90
WP
0.68
16.27
0.01 0.50 0.24 p90 0.68
15.79
0.07 0.30 0.38 p90
All
SP
103
0.77
12.46
0.38 0.66 0.01 p100
1.08
9.00
0.57 0.77 0.05 p95
WP
1.00
13.74
0.04 0.52 0.01 p95 0.87
10.38
0.15 0.63 0.13 p95
First
SP
103
1.00
12.46
0.38 0.66 0.01 p100
1.04
8.71
0.59 0.74 0.04 p95
WP
0.95
13.84
0.04 0.55 0.04 p95 0.91
10.33
0.07 0.63 0.08 p90
358
May 2016
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
