04 International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004
00 each pixel in the orthophotos into one of five classes: sand,
ng = 34 a lee sandy gravel, pebble, clean gravel, and cobble (Figure 4). Type
js pue 3 E 1 signatures, initially based on three colour bands, were derived
Sh mi TER from five of the fifteen classified sub-areas and located by
ng ground survey (Dataset 3). The true accuracy of the
ies classification was then assessed by comparing the classification
to od in the remaining sub-areas with their known bed material
^" categorisation. It is important to stress that these areas were
the determined on the ground but were not used during signature
ent derivation, which contrasts to the dubious practice of deriving
1020 "accuracy" statistics from the same areas used to create
signatures. Results are presented in the form of an accuracy or
ms
ea. 1020 1040 1060 1080
ple - : ;
ser
0 x
get m
ese
es" |
ese |
scd
Kr Figure 3- Orthophoto of test area- 1:5,000 imagery
ha :
ther
low 2.2 Initial photogrammetric processing
)ns. : ; ;
sin Once images were downloaded it was possible to use
ales conventional photogrammetric methods to derive colour ortho-
Meo photographs for the 120 x 80m test area. This was achieved
mes - using Erdas Imagine OrthoBase (©Erdas LLC), following an
t 24 "in-situ" calibration of the camera using an off-line self-
the calibrating bundle adjustment (Chandler et al, 2001).
but Automated DEM extraction methods were capable of generating aos
J. to a DEM, but orthophotos were derived using the more accurate Ne Ton toto 108 no
000 DEMs measured in the field. All available imagery was Figure 4- supervised classification (5 classes), 1:5,000
mes processed and full orthophoto coverage was achieved at the :
ique !-10,000 and 1:5,000 photo-scales (Figure 3). The central area contingency matrix in which the columns represent each of the
ther was derived at 1:3,000 and finally a small area within this was test areas (truth), whilst rows indicate the percentages of pixels
e of achieved at 1:1,000. classified into each of the 5 classes. The initial results were
disappointing (Table 1- average success rate 38%), with only
—— 3. RESULTS Cobble areas being identified with a high success rate (90%).
; ; Both Atkinson and Lewis (2000) and Lane (2001) recommend
3.1 Bed classification the addition of a “texture” layer to enhance spectral
Once orthophotos had been generated a conventional classification, An initial and imp le texture layer was derived
"supervised classification" procedure was used to categorise using a 3x3 variance convolution filter, which enabled localised
d variability in the image
Classified\Truth| Sand | Sand gravel | Clean gravel | Cobble Pebble to, be represented, This
1 simple addition
Sand 13 24 | 0 n/a improved accuracies
: Sand gravel 16 23 8 1 n/a radically (Table 2).
Clean gravel 25 25 27 9 n/a The average success
: rate improved overall
Cobble 20 8 62 90 n/a ta. 49%, with Sand
Pebbl 25 20 3 0 n/a exhibiting an accuracy
Table 1, 1:5,000- Percentages classified from three colour bands improvement of 26%.
Classified Truth| Sand Sand gravel | Clean gravel Cobble Pebble One additional test that
Sand 39 6 0 0 n/a was carried out was to
S : assess the significance
Sand gravel IT ae S 9 xi of the ten colour
Clean gravel 4 16 32 18 n/a bands compared with
Cobble 18 9 59 81 n/a simple grey-scale
Pebbl 28 26 I ne n/a image representation
ma Table 2- 1:5,000- Percentages classified from three colour bands + texture layer combined with texture.
1081
