Speckle Reduction

Normally it is not necessary to call a specific speckle-reduction routine first (or indeed at all). However, when required despeckle will provide a despeckled version of a given image without unnecessarily losing detail.

The despeckle routine

despeckle approximates a single or multiple image with a locally-constant (i.e. locally-flat) model of the surface cross-section. It fits this model using a variant of the simulated annealing optimisation method. The noise model used in the procedure is determined by the looks measure for the image. A higher value of looks measure implies that the data are less influenced by noise and are, therefore, more reliable estimates of the underlying surface.

The InfoPACK routines all assume the input image has uncorrelated speckle. If your image is oversampled so that it includes correlations then use the shrink utility to remove them; see Section 6.5 for a discussion of how to do this optimally.

The despeckle algorithm can be performed upon images with more than one image layer. Full Python script and function syntaxes for despeckle are given in section A.4.9.

The effects of the parameters of the speckle-reduction functions are described below.

The number of iterations

despeckle is iterative. The longer it runs, the smoother the resulting surface. Figure 4.13 shows the effect of smoothing a section of a DLR ESAR image for 10, 50 (the default value), and 250 iterations using despeckle with the looks measure set to 1 and other parameters set to the default values.

The looks measure parameter

Figure 4.14 shows the result of increasing the looks measure used by the routine despeckle. Fifty iterations were used so compare it with Figure 4.13. When the looks measure parameter is increased, the algorithm interprets the input data as being less affected by noise. This means that despeckle will be more likely to interpret fluctuations caused by speckle as genuine jumps in the cross-section. The output is less smooth as a result. We recommend that normally you let the routine estimate the looks measure for itself.

Figure 4.13: (a) Unprocessed SAR image (b) despeckled with looks measure = 1 and 10 iterations (c) 50 iterations (d) 250 iterations

Figure 4.13(a) Figure 4.13(b)

Figure 4.13(c) Figure 4.13(d)

Figure 4.14: Despeckling the image of figure 4.13(a) with 50 iterations and (a) looks measure = 2 (b) looks measure = 4

Figure 4.14(a) Figure 4.14(b)