Comparing TerraSAR-X and RADARSAT-2 polarimetric data for automated sea ice classification

• Main Authors: Ressel, Rudolf, Singha, Suman, Lehner, Susanne
• Format: Conference Object
• Language: unknown
• Published: 2015
• Subjects: Institut für Methodik der Fernerkundung; SAR-Signalverarbeitung; Baffin Bay; Baffin; Sea ice; Svalbard
• Online Access: https://elib.dlr.de/97082/

In the field of sea ice monitoring, single-polarimetric SAR images have been used for decades by means of classical image analysis methods such as segmentation and texture parameters, followed by some subsequent classification process, both supervised and unsuper-vised. Research into the potential of polarimetric features for sea ice classification has only recently gathered increased momentum with the advent of new SAR sensors with polarimetric capability. Dual-polarimetric and multi-polarimetric data offers the advantage of exploiting different polarimetric behavior of different ice types. Most attention into this direction has so far been devoted to dualpolarimetric data in C-band, most noteably through RADARSAT-2 data. We therefore strive to complement on the knowledge of polarimetric SAR imaging by conducting a comprehensive analysis of simultaneously acquired datasets in C-band (RADARSAT-2) and X-band (TerraSAR-X) over ice infested areas. First, we propose an array of polarimetric features (Pauli based, Lexicographic based), both for dual-polarimetric data and for multi-polarimetric data. Our analysis is conducted on four different instances of simultaneously acquired RADARSAT-2 quadpol Stripmap images and dualpol TS-X Stripmap images (Baffin Bay, Svalbard coastal waters). The actual dominant ice situation was gathered through ice expert assessment and offcial ice charts in the respective areas, as well as SMOS data. From this information we then chose training datasets and validation datasets. In order to quantify the information theoretical relevance and redundancy of the features, we employed the concept of mutual information on the extracted feature data. Based on this statistical evaluation we then ranked the features in terms of relevance and inspected their redundancy. The most useful features were selected and this subset was ingested into a pixel based neural network classifier. The output was evaluated in terms of classification accuracy and compared for both sensors.