Classification of Multi-Temporal Sar Images and Insar Coherence Images using Adaptive Neighborhood Model and Simulted Annealing Approch

5. Results and Discussion
Multi-temporal ERS-2 SAR images over the Mekong river delta region, Vietnam, acquired on May 5, June 9, and July 14, 1996, respectively, are considered to test the algorithm. Their color composite image (RGB) is shown in Fig.2(a) (Liew, et. al., 1998). Each image has been processed as follows (Liew, et. al., 1998): (1) lowpass filter with a 5*5 averaging window; (2) downsampling to 50m pixel size; (3) adaptive Wiener filter; (4) 3*3 median filter. The cluster centers are calculated by applying FCM and PCM algorithms to the preprocessed data.

Table 1. Cluster centers calculated by FCM algorithm

cluster	C₁	C₂	C₃	C₄	C₅	C₆
Image 1	140.207	99.7053	0.17724	118.303	129.642	121.705
Image 2	73.3753	89.6236	0.13534	111.192	126.289	149.934
Image 3	82.252	84.6451	0.11518	98.9408	123.145	48.6108

Table 2. Cluster centers calculated by PCM algorithm

cluster	C₁	C₂	C₃	C₄	C₅	C₆
Image 1	144.393	91.4003	0.00348	115.729	139.493	122.39
Image 2	70.6142	83.862	0.00295	110.388	134.797	147.843
Image 3	78.9161	78.9335	0.002914	103.892	128.79	51.5453

Table 1 and 2 show that the cluster centers calculated from the two algorithms are very similar. Comparing the individual values, the relative difference (|c^F_ji-c^P _ji|/ c^F_ji) is smaller than 7%. One exception is cluster 3 (c3), but it represents the same class in these two results, i. e., those pixels that have small values and change rarely. The close agreement of the cluster centers is probably due to the fact that the noise in the SAR images has been reduced effectively. Using the cluster centers calculated by FCM and the classification algorithm described above, we obtain the classification result shown in Fig. 2(b). False color represents different classes. Even though only 6 classes are used, the main characters of multi-temporal SAR images are highlighted. There are no mixed-up classes within the homogeneous regions, e. g., the regions represented by red and green. Meanwhile, a lot of fine structures, e. g., roads are preserved.

Fig. 2 (a) Color composite image of multi-temporal SAR images. (Red: May 5, Green: June 9, Blue: July 14, 1996)(Liew, et. al., 1998); (b) False color composite classified image.
In addition, we now apply this classification algorithm to a InSAR coherence image that is processed from two JERS single look complex SAR images. They were acquired on June 1 and October 11, 1996, covering a region of south Sumatra, Indonesia. Fig. 3(a) shows the InSAR coherence image. The bright region (high coherence) is bare land and the dark region (low coherence) is forest. If we classify the whole image into two classes, i. e., forest and bare land, the cluster centers can be determined interactively. The classification result is shown in Fig. 3(b). Despite the high noise in the coherence image, the bare land and forest can be delineated clearly. However, if the coherence image is classified only according to the cluster center without considering (neighborhood) texture information, the image can not be classified properly (shown in Fig. 3(c)), and the mixed-up of classes is very serious.

Fig. 3(a) InSAR coherence image. The bright pixels (high coherence) correspond to bare land and the dark pixels (low coherence) to forest; (b) InSAR coherence image classified into bare land (bright region) and forest (dark region); (c) InSAR coherence image classified without using (neighborhood) texture information.
6. Conclusion
A set of multi-temporal SAR images is considered to test the classification algorithm. Firstly, the cluster centers are calculated automatically by using fuzzy c-mean and possibility c-mean algorithms separately. The obtained cluster centers nearly agree with each other. Secondly, each pixel is classified into different classes by minimizing an energy function. This energy function contains cluster centers and the neighborhood information. The neighborhood is chosen based on the local homogeneity. Thus, the fine structures in the image are preserved, meanwhile, mixed-up classes in the homogenous regions are reduced effectively. Finally, this classification algorithm is applied to an InSAR coherence image. The whole image is classified into forest and bare land. A satisfactory result is obtained, even though the coherence image is very noisy.

7. Acknowledgments
The author would like to thank Dr. Liew and Ms. Chen, for providing SAR images after registration and speckle noise removal; and Prof. Nokayama at Tokyo University of Agriculture and Technology, for providing a sample of JERS raw data.

References

Cherkassky, V. and Mulier, F., 1998. Learning from data: concepts, theory and method. John Wiley and Sons, Inc.
Garzelli, A., 1999. Classification of polarimetric SAR images using adaptive neighborhood structures. Int. J. Remote Sensing, 20(8), pp. 1669-1675.
Liew, S. C., Kam, S. P., Tuong, T. P. Chen, P., Minh, V. Q., and Lim, H., 1998. Application of Multitemporal ERS-2 synthetic aperture radar in delineating rice cropping system in the Mekong river delta, Vietnam. IEEE Trans. Geosci. Remote sensing Sensing, 36(5), pp. 1412-1420.
Hegarat-Mascle, S. L. and Vidal-Madjar, D., 1996. Applications of simulated annealing to SAR image clustering and classification problems. Int. J. Remote Sensing, 17(9), pp. 1761-1776.
Wong, Y. F. and Posner, E. C., 1993. A new clustering algorithm applicable to multispectral and polarimetric SAR images. IEEE Trans. Geosci. and Remote Sensing, 31(3), pp. 634-644.

Page 3 of 3
| Previous |