Segmentation and labeling of polarimetric SAR data: can wavelets help?

In this paper we report about a novel approach to segmentation and thematic labeling of SAR polarimetric data. A pre-processing phase based on a wavelet frame that works as a differential operator generates piece-wise smooth approximations of the covariance matrix power term images. This step matches the signal characteristics with the requirements of well proven and computationally robust clustering algorithms of the Bayesian MAP, hard or soft labeling, and contextual type. Segments defined in this first phase are then used to estimate polarimetric quantities such as the Cloude's and Pottier's target decomposition parameters. The advantage over methods that use local statistics in a moving window is that the variance of the estimators decreases with the segment size, good accuracy can be obtained without sacrificing spatial resolution, and errors due to the signal discontinuities are avoided. The passage is then made from pixel based clustering to segment wise thematic classification. To the purpose reference feature vectors based on ground truth are needed, and segments are reassigned to the thematic labels. The per-segment feature vectors, including averages of polarimetric, radiometric quantities, or texture measures, can also be exploited to derive biophysical parameters and qualify additionally the classification categories. The proposed approach seems to be appealing because it can tackle under a unified framework resting on solid mathematical foundations - such as Bayesian inference and wavelet theory - different data sources (microwave and optical) and different thematic contexts.