Unsupervised detection and quantification of iceberg populations within sea ice from dual-polarisation SAR imagery

Accurate estimates of iceberg populations, disintegration rates and iceberg movements are essential to fully understand ice sheet contributions to sea level rise and freshwater and heat balances. Understanding and prediction of iceberg distributions is also of paramount importance for the safety of commercial and research shipping operations in polar seas. Despite their manifold implications the operational monitoring of icebergs remains challenging, largely due to difficulties in automating their detection at scale.  

Synthetic Aperture Radar (SAR) data from satellites, by virtue of its ability to penetrate cloud cover and strong sensitivity to the dielectric properties of the reflecting surface, has long been recognised as providing great potential for the identification of icebergs. Many existing studies have developed algorithms to exploit this data source 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 characteristics within and between SAR scenes that result from meteorological, geometric and instrumental differences. Further ambiguity arises when detecting icebergs in dense fields close to the calving front and in the presence of sea ice. For detection to be fully automated, therefore, adaptive iceberg detection algorithms are required, of which few currently exist. 

Here we propose an unsupervised classification procedure based on a recursive implementation of a Dirichlet Process Mixture Model that is robust to inter-scene variability and is capable of identifying icebergs even within complex environments containing mixtures of open water, sea ice and icebergs of various sizess. The method exploits freely available dual-polarisation Sentinel 1 EW imagery, allowing for wide spatial coverage at a high temporal density and providing scope for near-real-time monitoring.  It overcomes many of the limitations of existing approaches in terms of environments to which it may be applied as well as requirements for labelled training datasets or determination of scene-specific thresholds. Thus it provides an excellent basis for operational monitoring and tracking of iceberg populations at a continental scale to inform both scientific and navigational priorities.