Investigation of C-Band SAR Polarimetry for Mapping a High-Tidal Coastal Environment in Northern Canada

Abstract

Synthetic Aperture Radar (SAR) has been used in characterizing intertidal zones along northern Canadian coastlines. RADARSAT-2, with its full polarimetric information, has been considered for monitoring these vulnerable ecosystems and helping enhance the navigational safety of these waters. The RADARSAT Constellation Mission (RCM) will ensure data continuity with three identical SAR satellites orbiting together, providing superior revisit capabilities. The three satellites are equipped with multiple configurations, including single-polarization (HH, HV, VV), conventional (HH-HV, VV-VH, and HH-VV), hybrid (i.e., compact) dual polarization, and fully polarimetric (FP) modes. This study investigates the potential of the compact polarimetric (CP) mode for mapping an intertidal zone located at Tasiujaq village on the southwest shore of Ungava Bay, Quebec. Simulated RCM data were generated using FP RADARSAT-2 images collected over the study site in 2016. Commonly used tools for CP analysis include Raney m-delta classification and the hybrid dual polarizations RH-RV (where the transmitter is right-circular and the receivers are horizontal and vertical linear polarizations) and RR-RL (where the transmitter is right circular and the receivers are right-circular and left-circular polarizations). The potential of CP is compared with single, conventional dual-pol, and FP. The Freeman–Durden and Touzi discriminators are used for FP analysis. The random forest classifier is used as a classification approach due to its well-documented performance compared to other classifiers. The results suggest that the hybrid compact (RR-RL and RH-RV) dual polarizations provide encouraging separability capacities with overall accuracies of 61% and 60.7%, respectively, although they do not perform as well as conventional dual-pol HH-HV (64.4%). On the other hand, the CP polarimetric m-delta decomposition generated slightly less accurate classification results with an overall accuracy of approximately 62% compared to the FP Freeman–Durden (67.08%) and Touzi discriminators (71.1%).