GISdevelopment.net ---> Technology---> Image Processing

Information Fusion Based on Multiresolution and Multisensor SAR and Optical-Sensor Segmented Images

Wihartini1 and Aniati Murni2
1Faculty of Engineering, Budi Luhur University
Jl. Raya Ciledug, Jakarta, Indonesia
Fax (62 21) 5853752,
e-mail: wihartini@bl.ac.id

2Faculty of Computer Science, University of Indonesia
Kampus UI, Depok 16424, Indonesia
Fax (62 21) 7863415,
e-mail: aniati@cs.ui.ac.id


Abstract -
This paper presents an information fusion method based on multiresolution and multisensor Synthetic Aperture Radar (SAR) and Optical-sensor segmented images. The result of information fusion using single resolution and multisensor segmented images gives a thematic image with artifacts. This paper has proposed an information fusion method using single or multiresolution and single or multisensor segmented images. The multiresolution segmented images are obtained using the A Troust wavelet transform and the information fusion is done using a composition approach. The results show significant improvement, such as better region boundaries of linear features more homogeneous regions of low-frequency types of object.

I. Introduction
The use of SAR image to obtain the thematic image of an area becomes more important due to the failure of using the optical-sensor image that contains cloud cover. In general, the thematic images produced by a classifier using either SAR or optical-sensor data have 'salt and pepper' effect. Solaiman et al. [1] has used the Nagao filter and Sun et al. [2] has used probability relaxation technique to smooth the effect of 'salt and pepper'.

The à trous wavelet transformation gives an offer to filter the low-pass and high-pass components in an image into different scales [8]. In this way, the method provides a possibility to obtain the best thematic image. The wavelet-based segmentation and threshold technique provides a selection of multi-resolution segmented images that is suitable to the objects of interest. A pair of the most suitable segmented images is chosen from the SAR and optical-sensor wavelet-based segmented images and threshold technique. A reference map is used to do the task. Finally, the detail optixal-sensor segmented image is verified by the information of the detail SAR image.

This paper is organized as follows. Section 1 presents the introduction and Section 2 consists of the data and methodology. The following Section 3 presents the a trous wavelet transformation that is used for filtering and segmentation and thresholding process.. Section 4 presents and discusses the experimental results. This paper is closed by a concluding remark.

II. Data and Methodology

2.1 Site Area
The object area of the research is Central Kalimantan, located at longitude 113053'-114000' and latitude 2008'-2011'. The vegetation in this area consists of scattered trees, rice paddies, swamps and woodlands. The data to be used is a JERS-1 SAR L-band scene (path-row: 96/304) taken on 20th March 1997 and LANDSAT TM taken on 24 March 1997. Data set is processed with standard geo-coded image re-sampled to the UTM (Universal Transverse Mercator) projection by NASDA. Map-to-map registration is used by choosing 25 GCPs (Ground Control Points), using polynomial functions with accuracy of less than 1 pixel RMS (Root Mean Square) error, and resampled with 12.5 m resolution for each pixel.

2.2 Methodology
The process in this research is divided into two stages: pre-processing and intermediate processing. Pre-processing stage is doing registration and filtering, while intermediate processing stage is doing segmentation and object identification.

Image registration is done by map-to-map registration from Landsat image to SAR image. Subsequently, denoising process on both images is done with à trous wavelet transformation. The segmentation process is performed for each detail image of both sensor images. The detail image is obtained by à trous wavelet transformation up to fourth level. From the resulting detail image segmentation, object identification on SAR image is performed based on the Landsat detail segmented. The whole process can be seen on Figure 1.



Figure 1. Information fusion flowchart


III. À trous Wavelet Transform

3.1. À trous algorithm
A wavelet transformation for discrete data can be obtained by a version known as the à trous algorithm, which means by "hole". The à trous algorithm is a non-dyadic resolution; it is a redundant or stationary transformation because there is no decimation process. Wavelet transformation with à trous 1-dimensional algorithm will produces one set {wj} on scale j, which contains the information detail and a scalar product cj that contains image approximation information. Both wj and cj has the same pixel number with the original image.



3.2. Denoising
Denoising or noise reduction process is different from filtering. Filtering is done in frequency domain, that is, disposing the unwanted frequency area by thresholding where some information is disposed. But denoising is a process to dispose noise by pyramidal algorithm (Moorlet) or non-pyramidal (á trous ) algorithm.

The modeled image contains two components i.e. deterministic and stochastic, where the first is a signal while the second is a noise. There are two noise models: Gaussian additive noise model and non-Gaussian multiplicative noise model. Speckle is a non-Gaussian multiplicative noise. In SAR images, the noise is multiplicative, so components are multiplied. Homomorphic is applied to separate the deterministic and stochastic components, by taking the logarithm of the image: z = x v before processing, as follows :
log(z) = log(x) + log(v)             ...............................(9)
where : log(z) is the new image,
log(x) is the signal, and
log(v) is the additive noise.
 Since log(v) is an additive noise, then it can be treated in Gaussian model. Subsequently, for measuring the success level of speckle reducing process, it is done statistically. Generally, every image contains noise, so the wavelet coefficient also contains noise. The noise in the image follows the Gaussian or Poisson distribution, or the combination of both. In this study, multi-resolution support will be used to determine the noise significance.





3.3 Image segmentation
After SAR image and optical image has been denoised then à trous wavelet transform process until fourth level, next is segmentation process on fourth detail images (w1, w2, w3 and w4) on SAR image and optical image (figure 2.) SAR



Figure 2. Detail image w1, w2, w3, w4 from segmentation result SAR (a-d) and Landsat (e-h)


The image classification is done using the ERMAPPER software since ground truth is not available, thematic map and unsupervised classification is used with 7 predetermined classes (Table 1. and Table 2.) there are river(water), swamp, woodland, tropical grass, scattered tree, mangrove and rice paddy. Variant distribution that shown on Table 1 and Table 2 had been chosen as minimum and maximum value in threshold process on detail images w4 and w3 for each classes. In this research, fusion process only done on detail images w4 and w3 from SAR and Optical images.

Table 1. Unsupervised classification w4





Table 2. Unsupervised classification w3



4. Experimental Results and Discussions
From this segmentation will choose object that would be fusion. From 7 images, result of threshold process on SAR images and Optical images, choose an object that visually looks clearly or choose an object that have wide variant distribution on Table 1 and Table 2.

Fusion for detail image w4, the image taken from threshold process 6th , 7th classes of SAR images to 5th, 4th, 3rd, 2nd and 1st classes of optical images. The result of fusion shown on figure 3.



Figure 3. Fusion from w4, threshold SAR detail image to w4 threshold optical detail image


Fusion on detail image w3 done by using image from threshold process 6th , 7th classes of SAR images to 5th, 4th, 3rd, 2nd and 1st classes of optical image..The result shown on Figure 4..



Figure 4. Fusion from w3 threshold SAR detail image to w3 threshold optical detail image


V. Concluding Remarks
  1. Using threshold as chosen object would make fusion process easier.
  2. 2. The use of à trous wavelet transformation both in the denoising process and segmentation process can give several alternative segmented images.
  3. Composing or fusing the chosen two sensor segmented image can be done to obtain a reasonably better thematic image

References
  1. B. Solaiman, R. Koffi, M-C Mouchot and A.Hillion, An Information Fusion Method for Multispectral Image Classification Postprocessing, IEEE Transaction on geoscience and remote sensing, Vol. 36 No. 2, pp. 395-406, March, 1998.
  2. W. Sun, V.Heidt, P. Gong and G. Xu, Information Fusion for Rural Land-Use Classification With High-Resolution Satellite Imagery, IEEE Transaction on geoscience and remote sensing, Vol. 41 No. 4, pp. 883-890, April, 2003.
  3. R. Huber and L.V. Dutra, Classifier Combination and Feature Selection for Land-Cover Mapping from High-Resolution Airborne Dual-Band SAR Data, Vol. V, Part I, pp.370-375, Orlando, USA, July, 2000.
  4. T.M. Lillesand and R.W. Kiefer, Remote Sensing and Image Interpretation, John Wiley & Sons, Inc., 1994.
  5. R.M. Haralick, K. Shanmugam, and I. Dinstein. Textural features for image classification. IEEE Transactions on System, Man, and Cybernetics, 3(6): 610-621, November 1973.
  6. A.H. Schistad Solberg, A.K. Jain, and T. Taxt. Multisource classification of remotely sensed data: fusion of Landsat TM and SAR images. IEEE Transactions on Geoscience and Remote Sensing, 32(4):768-778, 1994.
  7. A. Murni, A.K. Jain, and J. Rais. A Framework for Multidate Multisensor Image Interpretation. In Proc. International Geoscience and Remote Sensing Symposium, pages 1851-1854, 1996.
  8. S.Sardy, Wihartini and S. Jatno, The Area Change Detection on Synthetic Aperture Radar Images by using Wavelet Transform, Proceeding of the IASTED International Conference, Marbella, Spain, September 4-7, 2001.
  9. Wihartini, Wiweka and Aniati Murni, SAR Image Classification using Information Method, Proceeding of the IJJS International Conference, Chiba University, Japan, October, 2004.
© GISdevelopment.net. All rights reserved.