International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 
  
scaled onto a 0-255 range, and displayed as a grey scale image. 
Darker pixels are spectrally "nearer" to their class centroid (in 
the sense of statistical distance), and are thus more likely to be 
classified correctly. On the other hand, pixels with higher 
distance values are spectrally further from the centroid of the 
class to which they were assigned, and are thus more likely to 
be misclassified. A threshold can be applied to the distance 
image to identify those pixels that are most likely to be 
misclassified. 
349 
21 
20 
1 
1 
   
z 
Table 2. Confusion or error matrix (source Vieira and Mather, 
2001). 
An alternative way of looking at the spatial distribution of the 
errors present in a classified image is by directly comparing 
thematic images with their respective ground truth maps. One 
of the products of this comparison should be a binary error 
image (Figure 1(b)) in which each point takes the value 0 
(correctly labelled) or 1 (erroneously labelled). By examining 
the spatial distribution of such pixels in Figure 1 we can make a 
number of observations. It is apparent that misclassified pixels 
are spatially correlated. These correlation effects are probably 
due to the presence of mixed pixels at field boundaries, to 
variation in the reflectance spectrum caused, most probably, by 
variations in soil type within a field, or to the effects of crop 
management practices such as the use of fertilisers. Spatial 
analysis measures (e.g., Join Count Statistics) could be used in 
order to determine whether these correlation effects are random 
or clustered in their spatial distribution. Looking at the spatial 
distribution of the remaining errors can help to refine the 
classification process. 
  
[1 Cones Plank 
WS Wong Pret 
  
  
  
  
(a) Distance Image (b) Error Image 
Figure 1. Spatial characterisation of classification errors 
using thematic image generated by Maximum Likelihood 
Classifier (385 by 385 pixels). 
982 
3.2.2 Visualising the Reliability: Any distance measure 
between the pixel and the mean pixel values (or prototypes) of 
each class can be used to compute a measure of reliability of a 
pixel's label. These measures of reliability could be then 
combined to the already assigned class label in order to 
generate a new thematic value for the pixel, which not only 
indicates the class to which the pixels was assigned but also the 
degree of accuracy achieved. A separate colour is assigned to 
each class. Within-class levels are also assigned separate shades 
of that colour, so that each class is represented by five shades of 
the given colour (see Figure 2(a)). This kind of representation 
allows the visual appreciation of the degree of accuracy of the 
classified crop. À contour representation of the reliabilities can 
also be used (Figure 2(b)). These types of representation help 
the user to identify portions of the thematic map that have 
reduced reliability. Although the final map may look uniform in 
its accuracy, it is actually a representative assemblage from 
several image processing procedures and refinements. It fs 
important for the user to known how these accuracies are 
spatially distributed in the image through a thematic reliability 
map. 
  
  
  
  
  
(a) Thematic Reliability (b) Reliability Contour 
Representation 
Figure 2. Representation of the reliabilities using a 
Maximum Likelihood classifier (385 by 385 pixels). 
4. RESULTS AND DISCUSSION 
A summary of the data used in this study is presented in Table 
A 
3. 
Table 3. Summary of the mean deviation, standard 
Points 
“ollected. 
3.5146 -1 -10.9780 2.5555 
Max. Errors 80.3581 | -85.4290' | -94.4135 
Between Features 
Generated 
(Equally spaced) 108.3339 
  
176.8237 
deviations used in the Trend and Accuracy analysis; 
senerated Point method; and Areal and equivalent 
Rectangles approach. 
4.1 Trend Analysis 
Trend analysis was carried out using the Student's 1 distribution 
and a confidence level of 90% (a = 0.10). The critical value 
obtained from statistical table for 28 control points is: 757 0.107 
1.703. Using the Equations (1) and (2) in order to compute 
estimated values of zy and 7x 
Internatic 
reU 
LANDS/ 
no trend 1 
CBERS: 
! 
no trend 1 
42 Acci 
Accuracy 
images ai 
computed 
and Table 
r 
dm. IR 
  
Table : 
accurac 
Adopting 
limit valu 
the Brazil 
two imag 
43 Gen 
The value 
method a 
for instan 
distances 
trend ana 
the E and 
The accur 
CBERS 
coordinat 
à given cc 
C 
Table 5 
analysi: 
The critic 
in Table 
1:100.00C 
class B, 
class. 
Trend an: 
direction 
images at