Despeckle and Geographical Feature Extraction in SAR Images by Wavelet Transform


Karunesh K Gupta
Instrumentation Group
B.I.T.S. Pilani
India, 333031
kgupta@bits-pilani.ac.in

Rajiv Gupta
Civil Group
B.I.T.S. Pilani
India, 333031

Abstract-
This paper presents an approach for despeckle Synthetic Aperture Radar (SAR) images. The features such as roads, airport runways, rivers and other ribbon-like shape structures can be detected by new wavelet function. In this framework, we firstly develop a despeckle technique based on wavelet transform achieving an acceptable compromise between speckle reduction and ribbon-like structure preserving. This is followed by recognition of ribbon-like shape structure based on new wavelet function proposed by Yuan Yan Tang (2003). Experimental results show that the proposed technique is capable of extracting the different width as well as different gray-levels roads.

I. Introduction
One of the most significant topics in pattern recognition is analysis of Ribbon-like shapes. It can be applied to recognize roads, airport runways, rivers etc in remote sensing images. Synthetic Aperture Radar (SAR) is a kind of high resolution imaging radar. It can generate images independent of time and weather condition. It also has the ability of penetrating through some depth of the earth or vegetation. SAR images, widely applied in many fields, such as agriculture, forestry, geology, hydrology etc., are inevitably accompanied with speckles due to the coherent nature of the imaging system. The presence of speckles in an image reduces the resolution of the image and the detectability of the geographical features. Geographical features such as road, airport runways and other ribbon-like shape structures detection belong to the category of line extraction or edge detection. Due to speckle noise, extraction of lines and edges from SAR images is difficult.

The statistics of the speckle was extensively studied in [2], with the conclusion that a SAR intensity can be modeled as multiplicative noise. Arsenault and April (1976) proved that for a logarithmically transformed SAR image, the speckle is approximately Gaussian additive noise [3]. Previous methods used to reduce speckle noise in images include speckle filters such as Median, Lee, Kuan, Frost, enhanced Frost, Gamma or Maximum a posterior (MAP). Wavelet technique has been applied to image processing for compression and denoising. Most of the wavelet-based speckle noise removal approaches can be described in the four steps [4]:

  1. Take a logarithmic transform on the SAR image (preconditioning).
  2. Apply the orthogonal DWT to obtain the wavelet coefficients [6].
  3. Choose a threshold ? corresponding to the noise varience and apply the soft thresholding.
  4. Apply the inverse orthogonal DWT and the exponential transform to reconstruct the denoised signal.

The multiplicative model is not accurate model because the ratios of the standard deviation to the signal value, the “coefficient of variation,” should be constant at every point in all the possible used image. The proposed method use Spatial-oriented tree (SOT) structure in wavelet domain and observed much better results. Despeckled image convolved with new wavelet function that has been constructed by Yuan Yan Tang to recognize ribbon-like structure.

II. Proposed Method
The method is divided in two parts. First the image is despeckled then ribbon like structure is recognized.

A.Speckle noise is removed by SOT structure in wavelet domain
  1. Set parameter a.
  2. Apply N-level Dyadic Wavelet transform for the image and set up uniform threshold.
  3. Find out wavelet transform modulus maxima (WTMM) map from wavelet coefficients.
  4. Multiply WTMM map coefficients by a.
  5. Apply uniform threshold in wavelet coefficient excluding WTMM map coefficients.
  6. Apply inverse Dyadic Wavelet transform and reconstruct image.

B.The ribbon-like structure is recognized by novel wavelet function. A novel wavelet function plays a key role in this scheme, which possesses three significant characteristics, namely, 1) the position of the local maximum moduli of the wavelet transform with respect to the Ribbon-like shape is independent of the gray-levels of the image, 2) When the appropriate scale of the wavelet transform is selected, the local maximum moduli of the wavelet transform of the Ribbon-like shape produce two new parallel contours, which are located symmetrically at two sides of the original one and have the same topological and geometric properties as that of the original shape, and 3) The distance between these two parallel contours equals to the scale of the wavelet transform, which is independent of the width of the shape.

III. Algorithm and Experiments
The wavelet transform is calculated discretely. The wavelet transform of image f (n, m) is:



TABLE 1








Step4. Select threshold T according background in the original image and proceed with threshold on the modulus image (if necessary).

Step5. For each point (x, y), its modulus of the wavelet transform is compared with one of its neighboring points along its gradient direction, if its modulus arrives at the maximum, it will recorded as the local modulus maximum flocmax.

Step6. For each point(x, y) with local maximum, search the point whose distance to (x, y) along the gradient direction is s. If it is a point with the local maximum, the central point is detected.

Step7. The primary skeleton is formed by all the points detected in Step 4.

Step8. Modify the above primary wavelet skeleton according to the foregoing modification program to obtain the final wavelet skeleton.

This technique is analyzed using MATLAB (version 6) software. Remote sensing image (single band) of 256 x 256 is taken as an input image. The Speckle noise is added with remote sensing image. Fig. 1 and Fig. 2 show the original image and noisy image respectively. Wavelet transform coefficients are calculated for three scales (N = 3). The result of apply wavelet despeckle is given in Fig. 3. Fig. 4 shows the output of ribbon-like structure extraction algorithm.


Fig1: Original Image



Fig2: Speckled Image



Fig3: Despeckled Image



Fig4: Road extracted from despeckled image


IV. Conclusions
A novel wavelet-based ribbon-like structure recognition method is proposed which is simple and efficient. The wavelet function’s scale is chosen according to width of road. With further refinement, it will be a practical method for automatic detection of ribbon-like structure in remote sensing images.

References

  1. Tang, Y.Y. and You, Xinge, “Skeletonization of Ribbon-like Shapes Based on a New Wavelet Function,” IEEE Trans. on PAMI, vol. 25, no. 9, pp. 1118-1133, June 2003.
  2. Goodman, J. W., “Some Fundamental Properties of speckle,” Journal of the optical society of America, vol 11, no 22, 1976.
  3. Arsenault, H. H. and April, G., “Properties of Speckle integrated with a finite aperture and logarithmically transformed,” Journal of the optical society of America, vol 11, no 22, pp. 1160-1168, 1976.
  4. Donoho, D. L., “Denoising by Soft-Thresholding,” IEEE Trans. on Information Theory, vol. 4, pp. 613-627, May 1995.
  5. Yang, L., Tang, Y.Y., and Suen, C.Y., “A width-invariant Property of Curve Based on Wavelet Transform with a Novel Wavelet Function,” IEEE Trans. on System, Man, and Cybernetics- part B: Cybernetics, vol. 33, no. 3, pp. 541-548, June2003.
  6. Mallat, S.G., “A theory for multiresolution signal decomposition: The wavelet representation”, IEEE Trans. on signal Processing, vol. 11, pp.674-693, July 1989.