the troposphere. This is illustrated in Figure 2, which shows
high elevation differences with respect to the
CDED
.
In general, scattering lengthens the signal’s path, resulting
in lower elevation values (Duda
et al.
, 2001). Particle
concentration, size, and altitude govern the influence of
scattering on the elevation, an effect which varies from the
sub-meter to meter scale (Carabajal
et al.
, 2006).
The variable elevation of the laser ground-footprint (~70
m) may influence the apparent ground elevation. Multiple
peaks in the return echo represent vertical variability within
the footprint. This can be an issue when none of the peaks
corresponds to the ground elevation (when dense canopy
prevents substantial pulse reflection from the ground, for
example). Also, multi-modal returns, such as the ones induced
by surface reflection from the top and bottom of a cliff (or
any large step structures) within the
GLAS
footprint are also
problematic. While both peaks provide a legitimate elevation
value, it is the altitude of the last peak that is provided
in
GLA14
(Brenner et al., 2011). Slopes can also generate
multiple peaks, resulting in inaccuracies because of the laser’s
footprint. For every percent of slope grade, an error >3 cm can
be expected due to pointing errors (Martin
et al.
, 2005).
Methodology
Selection of Indicators
An indicator defines the parametric conditions for which an
elevation point was considered erroneous due to the influence
of a given contamination source. Indicators were determined for
outliers, attitude quality, points affected by scattering and satu-
ration, noisy signals, the effects of land cover, and terrain slope.
To identify the outliers, a comparison was made between
GLA14
elevations and the
CDED
. As observed in Figure 2,
outliers are characterized by extreme elevation differences.
GLA14
points for which the elevation differences with
CDED
were higher than 50 m were flagged as outliers. This threshold
was used because it was the most restrictive found in the
literature (Toutin
et al
., 2013; Beaulieu and Clavet, 2009).
Attitude quality is evaluated by comparing the attitude from
the star tracker with the one generated from an independent
software solution (Schutz, 2001). An attitude quality flag is
then created based on their coherency. Possible values for this
flag are: Good, Warning, and Bad. Points for which the attitude
quality indicator presented a warning or points that were
flagged as bad were judged as potentially erroneous.
Elevation points that have been significantly affected
by atmospheric scattering will present a low energy
echo (Brenner
et al.
, 2007). The dynamic gain feature is
independently used to amplify the low energy returns
that satisfy the threshold criterion of that parameter.
Consequently, gain can be used as an indicator for detecting
GLA14
points affected by particulate scattering (and other
energy reducing phenomena). While Brenner
et al.
(2007)
used a gain value of 30 as a threshold for detecting scattering
contamination, a higher value of 50 was also used for this
purpose by the National Snow and Ice Data Center (
NSIDC
)
(
NSIDC
, 2014d). Since the aim was to create a database of
accurate reference points, the more restrictive value of 30
was used as a threshold on gain. A scattering flag, calculated
using
GLAS
’s 532 nm atmospheric channel, is also available in
GLA14
. However, all
GLA14
points within the
AOI
had been set
to invalid for this flag. Therefore, it was not considered in this
study and the gain threshold was the only indicator used to
identify the contamination due to scattering.
Points affected by saturation are detected using a
saturation threshold, which is dependent on the receiver’s
gain (Webster, 2013). The latter is used to compute a
saturation flag and eventually a correction for the affected
point. However, for
GLA14
data, it is suggested that flagged
points as well as those points deemed to be correctable, be
eliminated (
NSIDC
, 2014e). This advice was followed and
points contaminated by saturation were filtered out.
Reflectivity allowed for the identification of points affected
by both saturation and scattering. Indeed, reflectivity should
be high for points affected by saturation, while it should be
low for points affected by scattering (Zwally
et al.
, 2008).
According to this logic, and based on Zwally
et al
. (2008),
points with reflectivity values not included in the 0.05 to 0.95
range were removed from the dataset. The same strategy was
also used by Huang
et al
. (2013).
The signal to noise (
S/N
) ratio was used to identify the
measurements contaminated by instrument noise. It was
calculated from the maximum amplitude of the return echoes
and the standard deviation of the background noise (
NSIDC
,
2014c). Points with a
S/N
ratio below 10 were eliminated.
Variable elevations within the footprint can be detected by
monitoring the presence of multiple peaks. Echoes with more
than one detected peak were flagged as potentially erroneous
to prevent elevation disparities resulting from irregular
surface components (Gonzales
et al.
, 2010). The National
Digital Elevation Program (
NDEP
) suggests that points located
on a slope steeper than 20 percent grade should not be used
as reference points (
NDEP
, 2004). The slopes employed as
slope indicators were calculated from the
CDED
. Points located
on slopes steeper than 20 percent were thus discarded to
prevent planimetric errors from influencing the calculated
vertical accuracy (
NDEP
, 2004).
Filtering GLA14 Data
First, all data was converted to the same spatial reference. To
do so, all elevations from
GLA14
were translated to orthometric
height (
EGM96
), and the
CDED
planimetric reference system was
converted from
NAD83
to
WGS84
. Then, the selected indicators
were used to identify potentially erroneous elevation values,
as outlined in Figure 3. Finally, a quality analysis was per-
formed on both the rejected and retained points, to validate
the quality of the filtering method.
Also, since snow cover normally occurs from October to
April over the
AOI
, its impact on the vertical accuracy of the
CDED
was analyzed by disregarding all the elevations points acquired
during this period. The indicators were re-evaluated and com-
pared with the results that included the snow cover period.
Quality Analysis
The root mean square elevation error with a confidence level
of 95 percent (
RMSE95
) was used as the measure for vertical
accuracy (
NDEP
, 2004). The
RMSE95
was calculated using the
elevation difference between the
CDED
and
GLA14
data (
CDED
−
GLA14
). Since the outliers are characterized by high eleva-
tion differences (Figure 2), and thus will distort the
RMSE95
,
the reference value for the quality analysis was calculated
after the removal of the outliers. The gain in vertical accuracy
(reduction in
RMSE95
) was calculated from the difference in
the vertical accuracy of the points remaining after filtering
compared to the reference value. The
RMSE95
difference val-
ues (
RMSE95
) were computed for each indicator to provide an
evaluation of their individual performance;
RMSE95
were also
calculated for all indicators combined, to provide an indica-
tion of the performance of the method as a whole.
The second step of the quality analysis dealt with points
rejected by the indicators (an analysis performed in parallel
to the actual filtering protocol described above). The
RMSE95
of the rejected points was first calculated, to assess how
erroneous they actually were. Then, the proportion of rejected
points, with respect to the original number of points, was
calculated. This proportion helped weighting the impact
of the indicator, where removing few points will have little
impact on the overall accuracy, even with an elevated
RMSE95
.
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
September 2015
703
