vest of Melbourne, 
ort Phillip Bay. No 
hich minimises the 
' vegetation. This is 
| lag between data 
ce 1994). 
jatially differentiate 
image rectification 
stablished over an 
e three interpreters 
ame classification 
m likelihood in this 
ied prior to the 
vided the basis for 
assified areas from 
f particular interest 
lassified differently 
some quantitative 
je computed. Using 
ified pixels not in 
lygons and provide 
spective classes. 
ired by field survey 
; study investigates 
between the data 
Jes. Each image 
fied the image 
conditions: same 
using maximum 
(using nearest 
id all pixels to be 
ses. The change in 
the polygons in 
tects the source of 
tion (positional) or 
ether the positional 
eparable or not, 
nd measurement of 
dertaken. 
SSION 
le 1 indicates the 
r interpreters 2 and 
be significant, any 
ige of the spatial 
/er 900 polygons in 
class with some 
ed on the polygon 
is then possible to 
significant level of 
and shape can be 
| the class can be 
d on this approach 
Je indicate that, as 
ess certain but the 
1996 
detection of variability within these boundaries is able to 
be determined. However, this local spatial variability is 
less able to be detected with elongated class shapes, as 
sections of the same class boundary are too close to 
determine differences between uncertainties in the 
‘boundary’ polygons and ‘local’ polygons. As the 
interpreters are using the same image the uncertainties 
for each class fall within the data analysis stage in the 
proposed framework. 
  
  
  
  
Interpreter 1 (Control) 
Pixels in Class (Irrigated 
Pasture) 
Interpreter 2 Interpreter 3 
Irrigated Pasture 44768 51582 
Bare Ground 98 4552 
Saltmarsh 0 0 
Other 11698 430 
Total 56564 56564 
  
  
  
  
  
Extracted from Allan & Ellis (1996) 
Table 1 - First Stage (classification) 
Results shown in Tables 2 (a) & (b) represent a subset of 
the image used in the first stage of the case study. Only 
one class is shown in the tables as control although 
seven classes in all were classified. As previously 
mentioned, three interpreters rectified the same image 
from which they independently obtained a RMS of 
between 0.2 and 0.8 pixels. Resampling this image using 
nearest neighbour and cubic convolution respectively 
yielded consistent results. The same training data, 
independently determined by each interpreter, are used in 
both resampled images to classify the image into the 
seven classes as required. It appears from these results, 
and other classes used as control, that the magnitude 
and spatial distribution of error is less susceptible to 
differences obtained by interpreters in the 
rectification/resampling process. The degree to which this 
data processing stage contributes to overall class error is 
yet to be fully investigated. 
  
  
  
  
  
  
Interpreter 1 (Control) 
Pixels in Class (Irrigated 
Pasture) 
Interpreter 2 Interpreter 3 
Irrigated Pasture 45746 45205 
Bare Ground 58 29 
Saltmarsh 0 902 
Other 1364 1032 
Total 47168 47168 
  
  
  
Table 2 (a) - Second Stage (resampliing using nearest 
neighbour then classification) 
Internation 
  
  
  
  
  
  
Interpreter 1 (Control) 
Pixels in Class (Irrigated 
Pasture) 
Interpreter 2 Interpreter 3 
Irrigated Pasture 46212 45407 
Bare Ground 125 60 
Saltmarsh 3 821 
Other 828 880 
Total 47168 47168 
  
  
  
al Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
Table 2 (b) - Second Stage (resampliing using cubic 
convolution then classification) 
5. CONCLUSIONS 
This paper has attempted to present a framework to test 
error characteristics in a remote sensing environment. 
The utility of this approach is the spatial representation of 
error which enables quantitative error estimates to be 
determined. Future work will investigate the possibility of 
characterising error and uncertainty differently and the 
the differentiation of polygons in disagreement for 
inclusion in the error measurement process. 
Note: For this paper the terms error and uncertainty are 
considered to have the same meaning. 
6. ACKNOWLEDGEMENTS 
The authors wish to acknowledge the contribution made 
by colleagues in the Department of Land Information, 
RMIT University who willingly participated as image 
interpreters for this study. 
7. REFERENCES 
Allan, R.C., Medak, M., Taylor, L. & Ellis, G. 1996, 'A 
Test to Validate the Use of Stratified Random Sampling 
for Accuracy Assessment of a Forest Cover Map’, in 
Proceedings of 8th Australasian Remote Sensing 
Conference, Canberra, in print. 
Allan, R.C. & Ellis, G.P. 1996, ‘A Case Study to Quantify 
the Uncertainty of Source Errors in Remotely Sensed 
Data’, in Proceedings of 37th Australian Surveyors 
Congress, Perth, in print. 
Aronoff, S. 1982, 'Classification Accuracy: A User 
Approach’, Photogrammetric Engineering & Remote 
Sensing, vol. 48, pp. 1299-1307. 
Aronoff, S. 1989, Geographic Information Systems. A 
Management Perspective, WDL Publications, Canada. 
Cherrill, A. & McClean, C. 1995, ‘An Investigation of 
Uncertainty in Field Habitat Mapping and the Implications 
179 
  
  
  
  
   
   
   
   
   
   
   
  
  
   
   
   
  
   
    
   
   
   
  
   
   
  
    
   
   
  
   
   
      
    
   
   
   
   
   
   
   
   
   
   
   
  
  
  
   
  
    
  
   
   
   
   
   
   
   
  
  
   
   
    
  
PPT OS ENS