Nonlocal Multiscale Single Image Statistics From Sentinel-1 SAR Data for High Resolution Bitemporal Forest Wind Damage Detection

Change detection of synthetic aperture radar (SAR) data is a challenge for high-resolution applications. This study presents a new nonlocal averaging approach (STATSAR) to reduce the speckle of single Sentinel-1 SAR images and statistical parameters derived from the image. The similarity of SAR pixels is based on the statistics of
$3\times3$
window as represented by the mean, standard deviation, median, minimum, and maximum. K-means clustering is used to divide the SAR image in 30 similarity clusters. The nonlocal averaging is carried out within each cluster separately in magnitude order of the
$3\times3$
window averages. The nonlocal filtering is applicable not only to the original pixel backscattering values but also to statistical parameters, such as standard deviation. The statistical parameters to be filtered can represent any window size, according to the need of the application. The nonlocally averaged standard deviation derived in two spatial resolutions,
$3\times3$
and
$7\times7$
windows, are demonstrated here for improving the resolution in which the forest damages can be detected using the VH polarized backscattering spatial variation change.