combination of these characteristics. Further examination 
of these polygons may reveal that the error is positional 
(data acquisition or data input stage), attribute (data 
analysis stage) or no discrimination is possible. 
Progression through to the error quantification/spatial 
variability stage is then possible to determine overall 
class accuracy estimates. 
For this particular case study, each polygon in 
disagreement is created from an overlay of multiple 
independent realisations of the same classified image 
using remote sensing analysts. The size, shape, 
perimeter and spatial distribution of these polygons 
indicates whether the classes are positionally misaligned, 
the variations in class specification are due to different 
interpretation of pixel values (classification) or the 
interpreters are unable to differentiate mixed pixel effects. 
Polygons (clumps) in disagreement are those aggregated 
pixels that have been assigned different classes by 
independent interpreters. 
| 
ERROR PROPAGATION ! 
MODELLING 
| 
  
ERROR DETECTION & MEASUREMENT 
  
  
  
  
  
  
  
  
  
  
  
  
  
MEASURE POLYGON 
CHARACTERISTICS |» ERROR DISCRIMINATION 
\ (POSITIONAL / ATTRIBUTE) 
SAME PHENOMENA 
REALISATIONS ERROR QUANTIFICATION 
SPATIAL VARIABILITY 
= 
  
  
  
  
  
/ 
| 
| 
| 
  
{ 
ERROR SOURCE IDENTIFICATION 
  
  
  
  
  
  
  
  
  
  
  
£ DATA > DATA 5 DATA 
: ACQUISITION : | PROCESSING ANALYSIS | 
os DATA | ccc o DATAS HC DATAUSAGE &.- 
: CONVERSION : | OUTPUT : :INTERPRETATION: 
  
  
  
Adapted from Lunetta et al (1991) & Veregin (1989) 
Figure 1 
International Archives of Photogrammetry and Remote Sensing. Vol. XXXI, Part B3. Vienna 1996 
3. CASE STUDY 
The study area is situated south west of Melbourne, 
Australia on the western shore of Port Phillip Bay. No 
public access to the site is allowed which minimises the 
degree of disturbance to ground cover vegetation. This is 
important when considering the time lag between data 
capture, integration and validation (Race 1994). 
Landsat TM imagery was used to spatially differentiate 
land cover into seven classes. For image rectification 
twelve ground control points were established over an 
area of 9km by 9km. In the first stage three interpreters 
classified the image using the same classification 
technique (supervised using maximum likelihood in this 
instance). The image was rectified prior to the 
classification. These realisations provided the basis for 
determining the degree to which the classified areas from 
each interpreter were in agreement. Of particular interest 
for this study were those areas classified differently 
between interpreters to enable some quantitative 
measures of these disagreements to be computed. Using 
the GIS overlay function the classified pixels not in 
agreement were clumped to form polygons and provide 
quantitative estimates of error in the respective classes. 
Using the ground control points, acquired by field survey 
using GPS, the second stage of this study investigates 
the accumulation of source errors between the data 
processing to data analysis stages. Each image 
interpreter rectified and classified the image 
independently based upon three conditions: same 
classification technique (supervised using maximum 
likelihood classifier), resampling (using nearest 
neighbour and cubic convolution) and all pixels to be 
classified into one of the seven classes. The change in 
the geometrical characteristics of the polygons in 
disagreement, in some instances, detects the source of 
uncertainty either from the rectification (positional) or 
from the classification (attribute). Whether the positional 
and attribute uncertainties are separable or not, 
progression through to the detection and measurement of 
local spatial variability can then be undertaken. 
4. RESULTS AND DISCUSSION 
Using interpreter 1 as control, Table 1 indicates the 
disagreement in pixel classification for interpreters 2 and 
3. Whilst these differences appear to be significant, any 
further analysis requires a knowledge of the spatial 
distribution of error for each class. Over 900 polygons in 
disagreement were formed for this class with some 
polygons as small as one pixel. Based on the polygon 
characteristics and visual display, it is then possible to 
determine which polygons indicate a significant level of 
error. Threshold limits based on area and shape can be 
set and the location of uncertainty in the class can be 
examined. Allan & Ellis (1996) expand on this approach 
with tests for other classes. 
Preliminary results from the first stage indicate that, as 
expected, the class boundaries are less certain but the 
      
  
  
  
   
   
   
  
  
   
  
  
  
   
   
   
   
    
   
   
  
  
   
   
   
  
    
   
   
  
    
  
   
   
  
  
   
  
  
  
  
    
    
  
  
  
   
  
  
  
   
   
detection of 
be determir 
less able to 
sections of 
determine 
‘boundary’ 
interpreters 
for each cl 
proposed fr 
Irrigated | 
Bare Gi 
Saltm: 
Oth 
Tot 
  
Results sh« 
the image 
one class 
seven cla: 
mentioned, 
from whic 
between 0. 
nearest n« 
yielded cc 
independel 
both resar 
seven clas 
and other 
and spatie 
differences 
rectificatio 
data proce 
yet to be fi 
Irrigated 
Bare C 
Saltrr 
Ott 
  
To 
Table 2