The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B4. Beijing 2008
1353
ArcMap for interpretation and class labeling. Objects relating to
each signature were identified in the second-unsupervised
classified image and the classes were labeled through various
image interpretation techniques. Subsequently, final signatures
were used as the input signature file in a supervised
classification procedure to extract a semi-supervised classified
image (Fig. 3). Erdas Imagine Accuracy Assessment function
used for the analysis of the results of supervised and
unsupervised methods. For this matter, control points were
selected with fair distribution over the image and their identity
were recognized using all image data and vector maps. For each
clip of the image, at least 20 control points were used; bearing
in mind to choose at least one control point for each of the
object classes. These control point then were used in the
accuracy assessment process. We repeated accuracy assessment
procedure for all the classification results of MS and pan-
sharpened images. In aggregate, the accuracies of unsupervised
and supervised methods in Kappa index were calculated about
0.8 and 0.9, respectively. Although the accuracy of the
supervised method is better than the unsupervised one, there
were larger areas in the supervised results which were assigned
as unclassified due to the lack of comprehensive training
signatures.
Figure 1. Supervised classified image
Figure 2. Signature plot of unsupervised classified image
Figure 3. Unsupervised (semi-supervised) classified image
It becomes obvious that any classifier requiring high training
accuracy may not achieve good generalization capability (Ng et
al, 2007). That is the reason we have larger unclassified areas in
the supervised method’s results.
5.4 Fuzzy Object Extraction
For the fuzzy object extraction method, we produced files of
training sites for all tracks of the image. In order to minimize
redundant differences in homogeneous areas of the image, a
segmentation function by eCognition software applied on the
image tracks. After segmentation, training sites were used to
define object classes. At first, parent classes labeled, then,
different classes in each parent class were defined. At last, all of
image tracks were classified and converted to polygon vector
layers (Fig. 4). In some cases, there were many pixel sized
polygons in the extracted vector layers. To minimize the
number of small polygons, classification process repeated from
the segmentation step down to vectorization phase. As the last
method of image information extraction, Neural Network
classification implemented with the help of IDRISI. For this
method, training sites defined in IDRISI and the tags of related
object classes were assigned. The classification process repeated
with 1000 iteration to achieve the least RMS error.
Figure 4. Fuzzy classified image using eConition