International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
The overall average success rate was lower (44%) but accuracy 
was better than may have been expected and is of particular 
are many large collections of black and 
significance. There 
hy of gravelly river channels and 
white vertical aerial photograp 
significantly, these record changes that have taken place over 
many decades. The spectral qualities of these archives have 
been largely unused and the simple techniques described could 
be used to obtain informative long-term records. 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
i 
a 
e 
a 
d 
s 
a 
V 
W 
Figure 6 Percentage sand from aerial photos (1:5000) It 
methods. It had been hoped to derive quantitative data to d 
demonstrate the accuracy of the image based approach. ; 
However, the contrast in resolution prevents an objective and dt 
quantitative assessment being made. by 
The automated aerial classification procedure was carried out vs 
also using the 1:10,000 orthophoto. A very similar result was et 
achieved, and only a slight smoothing effect was apparent when T 
compared to the 1:5,000 percentage sand map (Figure 6). bc 
Figure 5, Percentage sand from ground truth ed 
eo 
int 
3.2 Percentage Sand au 
wi 
Percentage sand is increasingly recognised as a key parameter in m: 
defining hydraulic and transport characteristics in gravel-bed "m: 
rivers and the ability to generate just percentage sand from fro 
aerial imagery is significant. Oo! 
wh 
The Wolman data from 10 x 10m sampling units (Dataset 2) 
were manipulated to yield percentage sand values across the full 4.1 
120 x 80m test area (Figure 5). It must be remembered that the 
sampling effort involved in obtaining this map (for a relatively An 
small part of the river bed) was great, yet resolution is coarse cor 
and inevitably key facies boundaries are transgressed. The (20 
classified image derived using the 1:5,000 orthophotograph and pro 
signatures based on three colours bands and the texture layer seci 
was then used to create an alternative percentage sand map. fou 
This was achieved by plotting the proportion of pixels classified higl 
as sand within a five by five pixel moving window (Figure 6). part 
- oni Fieure 7 Percentage sand- ground and aerial The 
If figures 5 and 6 are compared there 1s clearly a significant g © 5 whi 
improvement in the information provided. Facies units in the emp 
image have shapes and orientations that closely resemble their 4. DISCUSSION dist 
true nature and are clearly not influenced by the artificial blocks be i 
created by the conventional sampling strategy. Moreover, the The classification accuracies achieved is this pilot project, are sand 
automated classification procedure broadly mirrors that not as high as many published in the remote sensing literature, 
achieved by the ground fieldwork in terms of the estimation of with the highest average accuracy achieved being just 49%. The 
percentage sand across the entire surface. There are perhaps three factors that need to be considered. It tests 
should be remembered that sample sub-areas used to create corre 
This is further demonstrated if the ground and aerial estimates signatures were not homogeneous, with many individual alter 
of percentages are superimposed (Figure 7), where it is clear particles being outside the size range of the field identified type- of 
that localised peaks are smoothed by the course ground based In such inhomogeneous areas, a per-pixel classification will Same 
class 
1082