Comparing the potential of dual-pol TerraSAR-X, Sentinel, and Radarsat data for automated, polarimetric sea ice classification

Kurzfassung

In contrast to SAR single-pol data, which allow only classical image analysis, SAR dual-pol imagery can be analyzed by means of complex polarimetry. Our work investigates the potential of different dual-pol configurations (co-pol, compact polarimetry) in different satellite SAR sensors (TerraSAR-X, Sentinel; Radarsat) for automatic sea ice classification. The first step of our analysis comprises the extraction of polarimetric features. To enrich the information content of image segments, second order statistics on these polarimetric features are additionally computed. The discriminative power and relevance of the different features are ranked by utilizing the concept of mutual information. Different selections of the most relevant features are then fed into a neural network classifier. We explore different network configurations for optimal classification results. Performance is compared for different selections of relevant features. In order to evaluate the generalizability of trained classifiers, data for classification is taken from various geographical regions (Svalbard, Kara Sea, Baffin Island Coast, Antarctic). The outcome for the different sensors is then also discussed in terms of reliability and applicability. The implemented dual-pol processing chain exhibits improved performance over classical single-pol texture based ice classification approaches and is well-suited for fully automated ice charting purposes in near real-time situations. The promising results we achieved for our single-pol based classification algorithm during field campaigns (Akademik Shokalskyi, Polarstern, Lance) can therefore also be expected for dual-pol data, complementing our portfolio of navigation assistance products.