International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Vol XXXV, Part B4. Istanbul 2004 Interne 
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
  
   
   
  
  
  
  
  
  
  
  
Y 
Parcels of Selected 
na Categories Data 
Y 
Combine Je 
Y 
| Final Classified Data ed 
Figure 3. The flow chart of the multi-temporal masking 
classification technique. 
  
  
  
  
  
In this study, to overcome the spectral confusion between the 
agricultural land cover classes using multi-temporal satellite 
images, a sequential multi-temporal crop masking procedure 
was utilized through the classification. After each classification 
step, those categories accurately classified were taken out from 
the classification process. 
The procedure was carried out as follows: The classified images 
of the original bands, thc PCA channels and the optimum bands 
from each date were evaluated to determine the accurately 
classified land cover classes. Those common classes with high 
accuracy in May and July classified images were identified and 
their training set were taken out from the classification of 
August imagery. For a particular class to be masked out from 
further classification process the accuracy threshold was 
selected as 80% and 90% for the producer's and user's 
accuracies respectively. The following crops, which fulfilled 
above accuracy requirement, were determined not to be 
included in the classification of August imagery: clover from 
classified May PCA image, residue and rice from classified July 
original image with regional masking applied, and sugar beet 
from classified July PCA image with regional masking applied 
  
  
  
  
  
  
  
  
  
  
  
RESULTS 
The accuracy assessment of a classified image is an important 
step as it indicates a measure to show how reliable is the 
information extracted from the remotely sensed data. To assess 
the classified images, the ground reference data were compared, 
parcel by parcel, with each of the three classified images for 
May, July and August, respectively. 
On May image, overall the classification accuracy of the 
original bands (ETM+ bands 1 to 5 and 7) was 88.9%. That was 
0.5% more accurate than PCA bands and 1.4% more accurate 
than the optimum bands. On July image, the number of classes 
was increased to twelve since the crops seeded in May were 
grown and therefore represented heterogeneity within the study 
area. As well, owing to phenological evolution of the 
vegetation, the spectral variation increased within the parcels. 
Thus, the number of spectral class as belonging to cach 
information class increased per information class to be 
extracted. In this sense it was logical to expect drop in overall 
accuracy. The overall accuracy of 66.8% was achieved from the 
classified original bands of July. The overall accuracies for 
classified optimum bands and the PCA bands was 62.5% and 
61.2%, respectively. After applying regional masking, no 
change was observed in the overall accuracy of the classified 
original bands of July, however, improvements were observed 
in the user’s accuracy of some classes such as cauliflower by 
%]15. 
July image 
Most of the land cover types that were present in the 
even 
were also observed in the August image. A total of el 
classes were classified from the August image. Most of the 
crops were in their late stages of the development. Therefore, 
their spectral responses were better representative of their 
1088 
that simply could not be achieved, otherwise, through analysis (Table 1). Therefore, these classes were masked out and were inherer 
conducted on single date imagery. not included within further classification process. introdu 
class : 
n are c 
Producer's User's aecurac 
May July August Image Classified Class Accuracy | Accuracy overall 
£L (9/0) (%) optimu 
classifi 
1 3 A 
; May PCA Channels | Clover 100.0 100.0 overall 
Preprocessing was 7 
July Originals Bands classifi 
: | À (with regional Residue 93.1 92.0 Applyit 
| Training Area Selection | masking) the reg 
July Original Bands August 
(with regional Rice 80.0 100.0 The im 
3 3 3 e 
masking) through 
Optimum Band Selection July PCA Channels classific 
(with regional Sugar beet 80.2 95.5 classific 
4 4 masking) was 10° 
| Segmentaion Using Aprior Crop Boundary Info. | ihe ori 
Table 1. The accuracy of the classes to be used in multi- misclass 
| | 3 temporal masking. confusic 
: a. N Mask analyses 
| Maximum likelihood Classifier | NS à ; a enr 
Selected Categories The original bands of the August image were re-classified using remotely 
only the training areas of unmasked classes. At each step of 
3 3 3 pixel-based classification, successively, the crops not cultivated Augu 
W Parcel-Based Analysis 4 MLC within the crop cultivation zones were excluded together with 
the crops classified on previous dates. The classification output Clas: 
3 3 of the August image was then combined with the classification Corn 
Parcel-Based s of the masked classes 1 "der tc ain final s 
Select Properly Classified Categories Analysis outputs of the masked classes in of der to obtain final summer Residue 
crop inventory of the study area (Figure 4). ; 
Tomato 
A 
Sugar be 
Clover 
Pasture 
Pepper 
Waterme 
Unc/Bso 
Rise 
Cauliflow 
Overall 
Table 2. 
As a sun 
August in 
is not log 
the May 
than the 
provided 
band sets. 
in the cla: 
the other | 
of the M. 
clover, wh 
the classif 
for July.