Enriching radar scatterer annotation towards SAR benchmark data creation

Abstract

SAR imagery has been successfully employed for various machine/deep learning applications. CNN based land use land cover classification, and RNN based time series modelling are two examples. It is reported that the unavailability of extensive SAR benchmark data limits the applicability of using SAR data for machine/deep learning applications and the quality of the result. To address this, we attempt to develop methods to enrich the annotation of radar scatterers in SAR images. Particularly, when SAR images have information on multi-polarimetric channels, and additional topographic measurements are available, the annotation can include not only geometric features, but also physical and land-use features of radar scatterers. This study 1) uses a standard time series InSAR approach to obtain geometric features such as geo-position dynamics of radar scatterers; 2) utilizes a Random Forest classifier to categorize physical features of radar scatterers including surface, low, high volume and double bounce scattering mechanisms; and 3) assigns land-use features to radar scatterers with the help of external topographic measurements. We demonstrated our methods by using thirty co-polarimetric SAR PAZ data, and TOP10NL topographic base map, covering the province of Friesland, the Netherlands. In the end these annotated radar scatterers can be in the registry of SAR benchmark dataset.