International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B7. Istanbul 2004 
  
  
The strategy of comparing posterior probabilities from multiple 
resolutions was tested in the case study. For this approach, the 
classifier is applied at each resolution to obtain the probability 
P(kli) for each pixel k as a member of class i (i=1, 2, ..., m pos- 
sible classes). The probabilities are then converted to a posteri- 
ori probabilities of class membership, which are assessed as the 
probability density of a case for a class relative to the sum of 
the densities (Jensen 1996). The a posteriori probability of a 
pixel k belonging to a class i, L(i [k), is determined by the fol- 
lowing equation (1): 
Li|k)= ares (1) 
> a,P(k|i) 
i=l 
where P(k/i) = the probability for a pixel k as a member of 
class 1, 
a; ^ a priori probablity of membership of class i, 
m = total number of classes. 
For each pixel, the a posteriori probabilities sum to 1.0. At each 
resolution, the highest a posteriori probability and its related 
class are output for each pixel. L;(i/k ) represents the maximum 
posteriori probablity of a pixel k belonging to class i at resolu- 
tion level /, L(i/k ) derived from all resolutions and & is as- 
signed to the class with the highest maximum a posteriori 
probability. Thus, & in class c, if and only if, 
L(c]k ) 2 Li(ik ), (2) 
Where i = 1, 2, 3, ...m possible classes, 
[= 4m, 8m, ... possible resolutions. 
The results were evaluated and analyzed based on classification 
accuracy for eight land use/cover classes. A total of 600 ran- 
domly selected samples were identified for the study area. The 
overall and individual Kappa coefficients (Jensen 1996) were 
reported for the study area for a series of classification maps to 
evaluate the agreement between the classification results and 
the reference data. 
To determine the difference between two kappa coefficients, 
the significance test proposed by Cohen (1960) for comparing 
two classification results was adopted. With this method, the 
difference between two Kappa coefficients resulting from two 
classifications was first obtained. The square root of the sum of 
the variances Vy between the two classifications was then cal- 
culated. A z-value is determined by dividing the difference by 
the square root. A z-value above 1.96 indicates that two classi- 
fication results are significantly different at the 0.95 confidence 
level. 
Table 2 presents the summarized results obtained from single- 
resolution and multi-resolution approaches based on their clas- 
sification accuracies in discriminating between eight land 
1190 
use/cover classes. The Kappa coefficients obtained from classi- 
fication using three multiple strategies are greater when com- 
pared to those from single resolution image input. Classification 
accuracy improvements are significant at the 0.95 confidence 
levels for the multi-resolution approach relative to comparable 
results from all single resolution classifications. 
  
  
  
  
  
  
  
Resolution Kappa 
4m 0.614 
8m 0.648 
12m 0.667 
16 m 0.694 
20m 0.681 
Multi-resolution | 0.755 
  
  
  
  
Table 2. Summary of classification accuracies derived from 
single-resolution and multi-resolution strategies. Accuracy is 
expressed as Kappa values. 
4. SUMMARY AND CONCLUSION 
One of the fundamental considerations when using remotely 
sensed data for land use/cover mapping is that of selecting ap- 
propriate spatial resolution(s). With the increased availability of 
very high resolution multi-spectral images spatial resolution 
variation will play an increasingly important role in the em- 
ployment of remotely sensed imagery. The correct application 
of image classification procedures for mapping land use/cover 
requires knowledge of certain spatial attributes of the data to 
determine the appropriate classification methodology and pa- 
rameters to use. In general, traditional single-resolution classifi- 
cation procedures are inadequate for understanding the effects 
of the chosen spatial resolution. They have difficulty discrimi- 
nating between land use/cover classes that have complex spec- 
tral/spatial features and patterns. Although a number of differ- 
ent approaches have been developed for classifying highly 
heterogeneous landscapes, current research focuses on contex- 
tual, knowledge-based, and segmentation routines using spatial 
and spectral information. Most of approaches developed mainly 
for Landsat TM and SPOT HRV images are often scene spe- 
cific and untested on high resolution images (1 m to 10 m). 
The multi-resolution framework proposed in this paper recog- 
nizes that image classification procedure should account for 
image spatial structure to minimize errors, and increase effi- 
ciency and information extraction from the classification proc- 
ess. Selection of the training scheme and classification decision 
rules should be guided by specification of the type of scene 
model (H- and L- resolution) and level of spatial variance rep- 
resented by the image to be classified. 
The technical basis of the multi-resolution framework and its 
potential advantages over commonly used single-resolution 
classification procedures were introduced and discussed. A va- 
riety of approaches may be used to generate multi-resolution 
image data sets. Different spatial analysis methods can provides 
the above information to allow resolution effects on individual 
classes examined. Different strategies can be used to incorpo- 
rate information from multiple resolutions. 
The case study illustrated the potential of multi-resolution clas- 
sification framework. Using a simulated multi-resolution data 
set and one multi-resolution strategy, it was demonstrated that 
multi-resolution classification approaches developed could sig- 
nificantly improve land use/cover classification accuracy when