Assessment of SAR speckle filters in the context of object-based image analysis

MORANDEIRA, NATALIA SOLEDAD; GRIMSON, RAFAEL; KANDUS, PATRICIA

Remote Sensing Letters

Taylor & Francis

Año: 2016 vol. 7 p. 150 - 159

2150-704X

The initial step in most object-based classification methodologies is the application of a segmentation algorithm to define objects. In the context of synthetic aperture radar (SAR) image analysis, the presence of speckle noise might hamper the segmentation quality. The aim of this study is to assess the segmentation performance of SAR images when no filter or differentfilters are applied before segmentation. In particular, the performance of the Mean-shift segmentation algorithm combined with different adaptive and non-adaptive filters is assessed based on both synthetic and natural SAR images. Studied filters include the non-adaptive Boxcar filter and four adaptive filters: the well-known Refined Leefilter and three recently proposed non-localfilters differing, in particular, in their dissimilarity criteria: the Hellinger and the Kullback?Leiblerfilters are based on stochastic distances, whereas the NL-SAR filter is based on the generalized likelihood ratio. Two measures were used for quality assessment:ρ-index and κ-index. Over-segmentation was assessed by theρ-index, the ratio of the resulting number of segments to the number of connected components of the ground-truth classes. The accuracy of the best possible classification given on the segmentation result was assessed with ground truth information by maximizing the κ-index. A Monte Carlo experiment conducted on synthetic images shows that the quality measures significantly differ for the applied filters. Our results indicate that the use of an adaptive filter improves the performance of the segmentation. In particular, the combination of the mean-shift segmentation algorithm with the NLSARfilter gives the best results and the resulting process is less sensitive to variations in the mean-shift operational parameters than when applying other filters or no filter. The results obtained may help improve the reliability of land-cover classification analyses based on an object-based approach on SAR data.