considered to improve coverage as they are able to maintain
fixed courses in windy conditions (Klemas 2015), but may be
susceptible to high humidity and/or rain which may restrict
sampling frequency. Additionally,
UAV
deployment may be
hampered in certain cases by local regulations, especially
near harbours. Other sensors may also provide complimentary
material—e.g., very-high resolution side-scan sonar which has
been successfully used to detect both intertidal and subtidal
L. conchilega
aggregations (Degraer
et al.
2008).
The average resolution of local relief models (i.e., effective
pixel size) was within the spatial resolution required to moni-
tor small fragmented biogenic tube aggregations (i.e., < 0.5 m)—
see Degraer
et al.
(2008), Hendrick and Foster-Smith (2006),
Rabaut, Vincx, and Degraer (2009).
RMSE
were within expected
values with models georeferenced by the
RTK GNSSs
displaying
smaller errors (0.09 ± 0.04 m) than those georeferenced with
the handheld
GNSS
receiver (1.30 ± 0.35 m). Employing hand-
held
GNSS
receivers limits applications of the used methods for
systematic time series due to the large RMSEs, which generate
displacement between maps of different dates that can result in
false change detection when using analytical methods (Hussain
et al.
2013). Average precision was high (i.e., very low displace-
ment) with the estimate of 0.01 m smaller than that recorded
by previous studies using similar equipment—i.e., 0.04 m by
Bryson
et al.
(2013) with
KAP
and handheld
GNSS
receiver. This
enabled accurate distance measuring from individual
LRMs
(Smith, Chandler, and Rose 2009). Lastly, artifacts related to
environmental characteristics were found on six
DTMs
and four
orthophoto mosaics, reinforcing the need to regard environ-
mental conditions during acquisition. These were likely associ-
ated to light conditions and water accumulation into puddles
on the intertidal prior to surveying, resulting in distortion dur-
ing relief reconstruction (Aber, Marzolff, and Ries 2010)—e.g.,
Marzolff and Poesen (2009) and Smith
et al.
(2009). In sum-
mary, we conclude that
KAP
is a suitable technique for low-cost
remote sensing, but suggest a conservative acquisition frequen-
cy to overcome weather-related challenges and/or mosaicking
across an appropriate window of time.
Lanice conchilega
Demographic Structure and Its Consequences to
Ecosystem Engineering
Maximum likelihood classification is a widely used algo-
rithm for image classification and its pre
studies obtained highly variable accurac
saltmarshes (Sanchez-Hernandez, Boyd,
21–60% in coral reefs (Mumby
et al.
199
submerged vegetation communities (Dekker, Brando, and
Anstee 2005), 21–23% in forests (Sanchez-Hernandez, Boyd,
and Foody 2007). Similarly, our classification accuracy (i.e.,
true presence percentage) was highly variable. Additionally,
agreement between the automated classification and visual
identification was extremely low, likely due to uncertain-
ties during class signature delineation as similarity between
signatures can result in inconsistent classifications (Mumby
et al.
1997) (see Figure 2). Although signature separability
was assessed visually prior to
MLC
, and training samples were
adjusted accordingly, additional analytical assessment of sig-
nature separability may further improve results (e.g., Dekker,
Brando, and Anstee 2005). Lastly, alternative classification
methods may be considered since these may present higher
accuracy in some instances. Sanchez-Hernandez, Boyd, and
Foody (2007) obtained significantly higher accuracy using
support vector machine (92.0%) than
MLC
(64.8%) to map
coastal saltmarsh habitats.
Classification accuracy also seemed related to population
status and weather conditions during sampling. The high-
est accuracies in identifying both the presence and absence
of
L. conchilega
were attained for the material from June
2014. This period was characterized as a period of high
L.
conchilega
density (approximately 11 000 individuals·m
−2
)
(Alves
et al.
2017) very widely distributed on the coast (Alves
2017). Whereas, the months with the lowest accuracies, i.e.,
April 2015 and August 2014 had low densities, approximately
700 ind.·m
−2
(Alves 2017) and approx. 8000 ind.·m
−2
(Alves
et
al.
2017). The sampling campaign on June 2014 was also char-
acterized by overcast conditions, whereas those in April 2015
and August 2014 were performed under sunnier conditions
(Alves 2017). Weather conditions affecting luminance in the
sampled photos may have impacted image segmentation as
luminance is known to affect it (Cheng
et al.
2001). As such,
alternative analyses that account for differences in luminance
may improve classification results. These are presented below
along other suggestions to improve the overall performance of
our proposed methodology.
Edge delineation using relative elevation as a selection
factor resulted in low concordance between the automati-
cally generated edges and hand delineation with high varia-
tion between dates. This indicates that relative elevation is
insufficient to delineate protruding
L. conchilega
aggregations.
Consequently, we were unable to successfully apply the scor-
ing system developed by Rabaut, Vincx, and Degraer (2009)
as only aggregations of obviously high value could be reliably
identified. Nevertheless, results also revealed a relationship
among false absence/presence percentages and relief complex-
ity wherein areas with well-developed
L. conchilega
aggrega-
tions displayed lesser mismatch between automatically and
manually delineated edges than areas with small and sparse
aggregations. This was dependent on window sizes used dur-
ing
MES
construction which influenced the position of gener-
ated edges, modifying its placement on the transition gradient.
Future improvements could include conditional selection
of polygons based on the image classification outputs (e.g., by
selecting polygons using
L. conchilega
pixel counts from pres-
ence/absence rasters). Changing how the reference elevation
surface is calculated may also improve detection. We used
a moving window to create mean elevation surfaces (
MESs
)
as references to convert
DTMs
into
LRMs
. The process created
biases in detection that may be minimised by using
DTMs
cre-
ated from the interpolation of elevation values extracted sole-
ly from points in bare sediment. Unfortunately, this method
n this study because we were unable to
ts outside of
L. conchilega
aggregations
rpolation. Further improvement may
different colour quantisation methods
(Sebe and Lew 2000). We employed Red, Green, Blue (
RGB
)
orthomosaics in our analyses. The
RGB
colour space is highly
correlated (Cheng
et al.
2001) because it is the output of three
additive colour primaries (Sebe and Lew 2000). A transforma-
tion into other colour spaces such as Hue, Saturation, Value
could improve classification accuracy by allowing intrinsic
characteristics such as hue and saturation to be analysed
separate from luminance (i.e., value) (Cheng
et al.
2001). High
hue resolution has been shown to improve accuracy (Sebe
and Lew 2000) and should be considered in future work.
Sampling ancillary data from the environment may also
improve detection. Surficial chl-a content is often associated
to aggregation size and engineering capacity (Rabaut, Vincx,
and Degraer 2009). It is also an indicator for abundance of
photosynthesising micro communities inside
L. conchilega
aggregations (Passarelli
et al.
2012). Sampling chl-a can be
achieved by sampling the near-infrared (
NIR
) part of the spec-
trum using small modifications to the camera (e.g., Bryson
et
al.
2013; Pauly and De Clerk 2011). However, recent research
at Boulogne shows that the engineering effect of
L. conchilega
varies seasonally (Alves
et al.
2017), resulting in seasonal
variation of surficial chl-a content inside aggregations (De
Smet
et al.
2015). Thus, although imaging the
NIR
spectrum
PHOTOGRAMMETRIC ENGINEERING & REMOTE SENSING
December 2019
903
