Geography Reference
In-Depth Information
et al., 2008). At a lower scale, automated plant recognition
has almost exclusively only been undertaken for terrestrial
vegetation species. Du et al. (2007) used 15 morphological
features of leaves to classify 20 species. Prior to IA,
chlorophyll
a
fluorescence induction curves had also been
reliably used for automated identification of terrestrial
plants (Keranen et al., 2003).
The characterisation of body ornementation (e.g. pig-
mentation) can help to recognise and classify individuals
into species. For example, shape differences among indi-
vidual patches on the frontoclypeus can provide valu-
able information for rapid species identification of 10
Hydropsyche
species (see Statzner and Mondy, 2009 for
visual interpretation of digitalised images).
Among freshwater vertebrates, automated species
recognition using IA has mainly been developed for
fish. Recognition of fish species using imagery, for
conservation (e.g. when they migrate using fishways)
or commercial purposes (e.g. edible species), has been
significantly enhanced over the last 20 years (see a
review in Zion et al., 2007). Castignolles et al. (1994),
achieved species classification in fishways by analysing
morphometric features on multiple images of each fish
as it swam across the passage. However, for real-time
applications (i.e. when fish are lined up close to each
other), multiple imaging tends to be impractical. At
the same time as this development, Hatch et al. (1994)
demonstrated that automated fish counting and specia-
tion in fishways could also be performed from digitised
sequences of video frames (the process determined the
correct location and identification of 67 out of 70 fish for
three salmonid species). Cadieux et al. (2000) used a set
of infrared diodes and sensors that generate silhouettes
as the fish swim between them. Once acquired, the data
can be sent periodically to a computer using a direct link,
a satellite link or a cellular phone. Cadieux et al. (2000)
calculated some moment-invariants, Fourier descriptors
of silhouette contours, and the geometric features
described by Castignolles et al. (1994). A majority vote
method (Xu et al., 1992) was used to classify images
of five fish species with an overall accuracy of 78%.
This system allows the operator to select the species of
interest according to the fauna of the specified river.
Tillett et al. (2000) segmented fish images by means
of a modified point distribution model (PDM) which
considered the strength of an edge and its proximity,
to attract landmarks to edges. They estimated salmon
length with an average accuracy of 95% when compared
to manual measurement. However, their procedure
required manual placement of the PDM in an initial
position close to the centre of the fish, and some images
(a)
(b)
Figure 17.4
Image of Lake Biwa sediment, containing almost
completely obscured microalgae specimens (a); corresponding
fluorescence image (b). Scale bar
100 Am. Reprinted from
Journal of Microbiological Methods
, 51, Walker et al.,
Fluorescence-assisted image analysis of freshwater microalgae,
pp. 149-162. Copyright 2002, with permission from Elsevier.
=
(Figure 17.4) and analysing microalgae in sediment
samples containing complex scenes. They quantitatively
measured 120 characteristics of each object detected
through fluorescence excitation, and used an optimised
subset of these characteristics for later automated analysis
and species classification. They succeeded in classifying
two genera of microalgae (
Anabaena spp
.and
Microcystis
spp
.) with accuracy higher than 97%.
Early
in situ
detection of algae species and the estima-
tion of their potential to cause algal blooms is also possible
from IA. For that purpose, an autonomous underwater
vehicle equipped with a submersible microscope, video
recording system and water quality monitoring sensors to
detect the spatial structure of
Uroglena americana
(caus-
ing freshwater 'red tide') was developed (Ishikawa et al.,
2005). Objects corresponding to the target species were
detected and analysed by extracting 130 statistical and
morphometrical features. The numbers of
U. Americana
objects per-unit-time were counted and combined with
the recorded vehicle route trajectory data. Subsequently,
colonies could be enumerated. In some cases (e.g. in
palynology, see Weller et al., 2006), self-organised maps
(a form of artificial neural networks), based on mor-
phological and textural IA features, were used for image
clustering of samples.
The discrimination of submerged macrophyte species
from optical remote sensing is possible at a large scale by
using appropriate spectral regions (see a review in Silva
et al., 2008). As suggested by Rowlinson et al. (1999), an
application of this method could be the detection of alien
species in riparian zones (e.g. prior to their removing).
However, water characteristics (such as turbidity and
depth), the presence of epiphytes and physiological status
of vegetation can be a source of variation in plant spectral
signatures, and this variability can lead to poor results
from simple automated classification procedures (Silva
Search WWH ::
Custom Search