calculation. Many authors try to reduce the effects of radar intensity variation by application 
of standard or special filter techniques. 
We advocate another approach. We suspect that, for extended objects, the signal variation 
which is inherent to each locally limited radar subimage, is not only dependent on system noise 
and speckle effects, but also reflects, to some extent, local structures of the backscattering 
surface. The application of any smoothing filter would reduce or eliminate this influence which 
we want to exploit. Hence, we will not apply smoothing filters, but aim to model the local 
radar signal variation and derive a texture measure thereof. Quite evidently, this approach 
can only be applied to area objects; prominent point targets, either isolated or in regular or 
irregular formations, result in strong, concentrated intensity peaks in the image and have to be 
discarded. The basic idea can already be found in [Ebert 87]. The following text summarizes 
the proposed processing steps. 
• Detection of local intensity extrema: 
We suppose that the radar signal backscattered from a specific area object is composed 
of a certain distribution of local twodimensional ’peaks’ and ’valleys’ of characteristic 
size and shape. A local extreme, peak or valley, is defined at pixel p t j, if the intensities 
of its 8 neighbour pixels are lower or higher than that of the center pixel. Coordinates 
of extrema are compiled in two separate lists for peaks and valleys, respectively. 
• Description of the form of local extrema: 
In real SAR imagery the form of local extrema, peak or valley, is anything else but an 
ideal symmetric twodimensional analytic function. Nevertheless, we are able to define 
a mean diameter at the basis of each extreme, and measure the dynamic range of an 
extreme which extends from the respective intensity level to the top or bottom of that 
extreme. In order to effect a higher sensitivity of these measurements, we resample a 
window of 8 x 8 pixels, centered to the coordinates of an extreme, into a submatrix of 
64 x 64 (sub-)pixels via Fourier transform, addition of the necessary number of ZERO 
spectral components, and inverse Fourier transform. 
• Calculation of the distribution of the local extrema: 
Two parameters are calculated to describe the local distribution of extrema, namely 
the distances between each extreme and its nearest neighbours, and the local density of 
extrema, i.e. the number of extrema per area unit. Calculation of distances is performed 
after the determination of the complete network of minimum size triangles connecting 
3 extrema each. These parameters are evaluated separately for both, peaks and valleys. 
Thus, we will assemble, separately for peaks and valleys, 4 lists of measurements: 
1. size of an extreme, 
2. dynamic range of an extreme, 
3. distances between an extreme and its nearest neighbours, 
4. local density of extrema. 
Note, that these measurements are not related to the pixels of the image, but to the coor 
dinates of the extrema. The number of extrema is within the range of 1/10 to 1/5 of the 
number of pixels for most SAR images. These features, together with standard features and 
non-image related context, will then be used for segmentation and classification of the image 
contents. 
697