td
re
re,
%.
It
ite
1al
DC.
All
International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXX V, Part B7. Istanbul 2004
. : greyscale) and used
Classified Truth Sand Sand gravel | Clean gravel Cobble Pebble e as ian additional
Sand +2 3 0 0 n/a layer to the three
Sand gravel 10 45 5 0 n/a colour bands and the
Clean gravel 5 21 x 27 n/a (ERDE layer used
prior. Results
Cobble 8 9 52 73 n/a indicated only a
Pebbl 35 22 0 0 n/a modest
Table 3 Percentages classified from three colour bands, texture 4- "similarity index"
inevitably yield a range of particle types and hence, during
accuracy assessment many pixels will. be identified as
erroneous, when they actually represent valid particles. In
addition, it should be recognised that true accuracies have been
derived by comparing classified data with test areas not used for
signature generation. Finally, the imagery was acquired in poor
and overcast lighting conditions during light drizzle. In brighter
weather, improved results may be obtainable: at least conditions
were typical of mountainous environments!
It is clear that a supervised classification based upon the three
colour bands is inadequate, particularly for a five-fold
classification. A three-fold classification into just sand, pebble
and cobble was more successful (56%) but the more ambitious
target was pursued. The classification is significantly enhanced
by adding a “texture” layer, which implies that the spatial
variability of signal response, rather than simply the colour
characteristics of the pixels is important. An interesting variable
in this context is the lithotype mix represented in the sediments
with colour variation between clasts being important. It would
be necessary to examine other field sites before this influence
can be examined fully. The criticality of texture was expected
given the relative roughness and hence shading-induced
intensity variability of different bed material textures. Other
authors have noted the significance of texture in related work,
with Atkinson and Lewis (2000) reviewing "geostatistical"
methods for classification and describing various alternative
measures of "texture". These range from the statistics derived
from the simple variance (Section 3. I) through to spatial auto-
correlation (Section 4.1) and semi-variograms (Section 4.2)
Which were investigated in this study briefly.
4.1 Texture via spatial auto-correlation
An additional avenue explored was the use of spatial auto-
correlation to assist in the bed classification procedure. Rubin
(2004) used spatial auto-correlation to derive correlation
profiles between an original greyscale image of borehole-sand
sections and the same image displaced by one, two, three and
four, etc. pixels. The basic idea being that large particles exhibit
high correlations for initial displacements, while. smaller
particles generate low correlations for all pixel displacements.
These basic differences generate contrasting correlation profiles
which then vary spatially. Rubin (2004) then develops an
empirical calibration function between measured sand grain size
distributions and the correlation profiles that, it is claimed, can
be used to identify sand grain size distributions for borehole-
sand sections.
The basic correlation algorithm was implemented in Matlab but
tests revealed some concerns, particularly whether spatial auto-
correlation can provide a good basis for measuring texture. An
alternative approach was developed based upon the reciprocal
of the summed variances between a 5x5 image patch and the
same patch displaced by one pixel. This was adopted for bed
classification using the 1:5000 orthophotos (converted to
improvement (Table
3. average success
rate 51%), implying
that very little new information was being added to the
classification process and that the original texture layer was
sufficient. :
4.2 Texture via the semi-variogram
Another approach investigated was the use of semi-variograms
to perhaps provide an alternative measure of “texture”. Work
conducted by Atkinson & Lewis (2000) and particularly Chica-
Olmo and Abarca-Hernandez (2000) advocated this
geostatistical method most strongly. Their field of application
was quite different, and involved the classification of
Quaternary deposits using Landsat 5 imagery (Chica-Olmo &
Abarca-Hernandez, 2000). Their results indicated an accuracy
improvement of approximately 20% and the approach appeared
promising. :
|—e— LargeParticles |
£— SmallParticles |
Sand
Semi-variance
4 6 8 1 2 14 16 1
Lag (pixels) where: pixel =8mm 9
Figure 8- semi-variogram, 1:1,000, (NS)
To investigate the technique semi-variograms were computed
for three of the sub-areas (Dataset 3) labelled in the field as
Sand, Gravel and Cobble. (Figure 8). This plot certainly
suggests that production of a semi-varidgram could provide
useful information for bed classification, with contrasting
profiles being generated for the differing bed materials.
500 m mt tete rte tort de si - ;
| . —e— LargeParticles
| a 400 ^. —8— Small Particles
| o
|.$300 Sand
1.2 200
TE
| 9 199
0
2 4 6 8 10 12 14 16 18
o
Lag distance (pixels) where: 1 pixel 8mm
Figure 9- semi-variogram, 1:1,000, ( EW)
However, Figure 8 represents a North-South section through the
semi-variogram surface, whilst Figure 9 represents an East-West
section through the same surface. In this case there is not such a
clear distinction in the profiles, particularly between the Gravel
and Cobble sediment types. It is clear that semi-variogram
direction is a critical controlling parameter with such an
1083
