The effect of speckle filtering on SAR texture discrimination (original) (raw)

1996, Simpósio Brasileiro de …

In tropical ecology studies, forest classification is a key issue.Although there is no widely accepted forest classification criterion, it is recognized that texture is an important factor to discriminate forest types and other land cover, particularly when using radar images. Synthetic Aperture Radar (SARA) images, however, are contaminated by a multiplicative noise, known as speckle, wich disturbs the texture identification. Several filters have been proposed to atenuate this kind of noise, but the effect of these filters on texture is not well known. In this paper, textures are modelled by two-dimensional autorregressive (AR-2D) models. These models are estimated for each one of the samples of SAR textures before and after speckle filtering. Eight samples of primary forest and seven samples of non-forest (pasture and agricultural crops) were collect from SAREX data (Cband, HH polarization, 6m resolution, 6 looks) in the Tapajós National Forest (Flona) region in Pará state, Brazil. All these samples were filtered by a 3 x 3 Box filter and a 3 x 3 Frost filter. Euclidean distances were computed between the model coefficient vector of the samples and the average coefficient vector for the two classes (defined here as the class vectors) for the unfiltered and filtered cases separately. For all cases the coefficient vectors formed two separated clusters, corresponding to each one of the classes, in a non-linear mapping of the coefficient space. The conclusions are: AR modelling is an effective method to idendify and discriminate radar texture. The discriminatory power, however, is higher using the unfiltered channels than when the simple Box and the Frost filter are used.