Satellite Images Unsupervised Classification Using Two Methods Fast Otsu and K-means

: Two unsupervised classifiers for optimum multithreshold are presented; fast Otsu and k-means. The unparametric methods produce an efficient procedure to separate the regions (classes) by select optimum levels, either on the gray levels of image histogram (as Otsu classifier), or on the gray levels of image intensities(as k-mean classifier), which are represent threshold values of the classes. In order to compare between the experimental results of these classifiers, the computation time is recorded and the needed iterations for k-means classifier to converge with optimum classes centers. The variation in the recorded computation time for k-means classifier is discussed.


Introduction:
Satellite image classification is one of the final goals in the remote sensing; where the relationship between pixels is investigated in order to cluster the pixels into certain classes.In some cases, it is useful to have the computer sort out which pixels have similar characteristics (e.g., spectra) rather than to try to force the pixels into a class based on our culturally driven sense of their similarities.This is done using an unsupervised classifier.
The resultant classes will be indicative of the natural spectral clusters in the data.They may or may not correspond to land cover or material classes as we normally think of them.Because of this limitation, unsupervised classification is often used as a preprocessor for other algorithms.For example, when attempting to locate training regions for supervised classification, it is often useful to know where natural spectral groupings occur [1].
Two simple classifiers are usually used; the k-means or iterative self-organizing data (iso-data) [2] and histogram multi-threshold using Otsu algorithm [3], in which the histogram grouped into panes depending on some statistical properties, by using the kmean algorithm the threshold between certain two panes are affected with tolerance value in the changing in the class's centers.

Sample of Test Sceme:
The scene that used in this study is Al-Ramadi city, which located in Al-Anbar provenience of the Iraq region, and it covers (581.72)km 2 west of Baghdad.Geographic location of AL-Ramadi region had shown in figure 1

Method:
Since the Otsu algorithm for multi-threshold histogram is time consuming because it is need huge computation, therefore the fast algorithm that proposed by Liao [4] is adopted and applied on the histogram's gray levels of Al-Ramadi image to find the threshold values of the classes, while the algorithm that presented by Seber [2] for the k-means algorithm is adopted and applied on the test image intensity values to find the threshold values between classes.Although the both classifiers are cluster the image into classes, but the main difference between them is; in the fast Otsu classifier it determine the best threshold values between classes (i.e. the maximum and the minimum value for each class) while the k-means classifier after grouping the image into classes it present the centers of the classes, therefore in order to compare the results of k-means with the results of fast Otsu algorithm, the maximum and the minimum values of the classes that obtained using k-means classifier should be calculated since the classes centers is not necessary to be in the middle of each class range.

Results and Discussions:
In order to compare the two classifiers, Al-Ramadi image which has 827×866 pixels with 256 gray level intensities figure 2-a is used.The threshold values for different classes are calculated using two methods and the computation time are recorded (as illustrated in table 1), using 64-bit computer platform of core 2 Due 2.2GHz processer and using MATLAB Ver.2009a.
It is obvious that the two methods succeed to calculate the optimum values for threshold the images gray level intensities to reach the same best threshold values to separate the required classes, with very close values (maximum difference is1 gray level) as illustrated in table 1. Figure 2-(c, d, e, f) shows the classified test image of Al-Ramadi into four classes using fast Otsu method.
The main difference between the two methods is required computation time; where it is clearly that the k-means algorithm needs huge computation time comparing with the fast Otsu algorithm.Another feature can be noticed in addition to huge computation time which required for the k-means classifier is the variation in the recorded computation time and the number of iterations that needed to converge with the best threshold gray level intensities, this is due to the value of the difference between the selected initial centers values with the optimum centers values of the classes and as shown in table 2. To select initial center values, the range of the gray levels intensities are divided into equal linear difference and selected as initial class centers to generate the start point for the k-means algorithm which is not necessary concede with the optimum class centers, the number of required iterations to reach the optimum class centers has forward relationship with the differences between the initial and the optimum class centers.

Figure 1
Figure 1 Test Image of Al-Ramadi Depicted itsLocation in Iraq

Fig. 3
Fig. 3 Computation Time Curves Counted using (a) Fast Otsu, (b) K-Means Classifiers a b . The available scene was TM exposure at March 04, 1990, illustrated in figure 2-a, the upper left corner is 206 lat.33° 27' 21.44" N, long.43° 29' 17.91" E and the lower right corner is lat.33° 14' 14.05" N, long., 43° 44' 36.64"E. This region represents Alluvial Plain; the hot desert climate prevails in the sedimentary plain and the western plateau.It contains some vegetation cover, many soil and rock erosion noticed.