Unsupervised machine learning detection of iceberg populations within sea ice from dual-polarisation SAR imagery

Accurate quantification of iceberg populations is essential to inform estimates of Southern Ocean freshwater
and heat balances as well as shipping hazards. The automated operational monitoring of icebergs remains
challenging, largely due to a lack of generality in existing approaches. Previous efforts to map icebergs have
often exploited synthetic aperture radar (SAR) data but the majority are designed for open water situations,
require significant operator input, and are susceptible to the substantial spatial and temporal variability in
backscatter that characterises SAR time-series. We propose an adaptive unsupervised classification procedure
based on Sentinel 1 SAR data and a recursive Dirichlet Process implementation of Bayesian Gaussian Mixture
Model. The approach is robust to inter-scene variability and can identify icebergs even within complex
environments containing mixtures of open water, sea ice and icebergs of various sizes. For the study area
in the Amundsen Sea Embayment, close to the calving front of Thwaites Glacier, our classifier achieved a
mean pixel-wise F1 score against manual iceberg delineations from the SAR scenes of 0.960 ± 0.018 with a
corresponding object-level F1 score of 0.729 ± 0.086. The method provides an excellent basis for estimation of
total near-shore iceberg populations and has inherent potential for scalability that other approaches lack.