Semi-Automated Semantic Annotation of Big Archives of High Resolution SAR Images

Abstract

We demonstrate how to achieve a semi-automated and rapid semantic annotation of high resolution SAR image patches in the case of big image archives. To this end, we start with an already existing annotated satellite image data set and a validated multi-level annotation scheme aimed at semi-automated land cover and urban scene labeling. Our goals are mainly achieved by pre-defined patch cutting, feature extraction from image patches, and a cascaded active learning approach where unnecessary processing steps are skipped. This concept allows us to adapt the classification/annotation quality of a knowledge discovery in databases system to the actual user requirements. In addition, an easily understandable visualization of the retrieved results can be obtained by interactive graphical data mining, by the generation of map products, or by data analytics tools providing mainly statistical results and graphics.