Combined object detection and physical model-inversion für PolINSAR-Bilddaten
This proposal concerns research intended to demonstrate the merits of combining object detection and physical parameter estimation in the automatic analysis of Synthetic Aperture Radar (SAR) data. Object detection refers to the localization of instances of various object categories, such as roads, buildings or any other type of land cover, in SAR images. Physical parameter estimation, on the other hand, yields a salient description of a given instance in terms of category specific information such as forest biomass, soil moisture or building height by applying a given model to SAR observables. Both kinds of analysis have a long but largely independent tradition in remote sensing research. Object detection relies crucially on the extraction of meaningful features from given observables to resolve ambiguities that complicate the decision process; physical object parameter estimates constitute valuable items of information in this context. The estimation of these parameters, at the same time, is category specific and clearly profits from the availability of accurate detection hypotheses. The proposed research focuses on exploiting precisely this mutual dependence to obtain a feedback cycle in which both object detection performance and parameter estimation accuracy can be improved.