Land Cover Semantic Annotation Derived from High-Resolution SAR Images

Users of remote sensing images analyzing land cover characteristics are very much interested in classification schemes that define a consistent set of target categories. Up to now, a number of established classification schemes are mainly being used by interpreters of medium-resolution optical satellite images focusing on large-scale land cover. In contrast, we concentrate in this publication on the definition of a new classification scheme for high-resolution synthetic aperture radar (SAR) images that are mostly taken over built-up areas. Here, we can see many small details of buildings, industrial facilities, and infrastructure that have to be classified. However, the appearance of details in high-resolution SAR images is often difficult to understand for human observers, and, therefore, calls for an automated semantic annotation of the target objects that has to follow a number of specific scientific guidelines. We demonstrate that a selection of representative SAR images with subsequent feature extraction and relevance feedback classification during the generation of a classification scheme leads to a reliable definition of a new high-resolution multi-level SAR image classification scheme that can be applied globally for semantic annotation in an automated chain.