An adaptive machine learning approach to improve automatic iceberg detection from SAR images

Iceberg distribution, dispersion and melting patterns are fundamental aspects in the balance of heat and freshwater in the Southern Ocean; yet these features are not fully understood. This lack of understanding is, in part, due to the difficulties in accurately identifying icebergs in different environmental conditions. To improve the understanding, reliable iceberg detection tools are necessary to achieve a detailed picture of iceberg drift and disintegration patterns, an thus to gain further information on the freshwater input into the Southern Ocean. Here, we present an accurate automatic large-scale iceberg detection method using an alternative machine learning architecture applied to high resolution Synthetic Aperture Radar (SAR) images. Our method is based on the concept of adaptability and focuses on improving the performance of identifying icebergs in ambiguous environmental contexts with wide radiometric, textural, size and shape variability. The fundamentals of the method are centred on superpixel segmentation, ensemble learning and incremental learning. The method is applied to a dataset containing 586 ENVISAT Advanced SAR images acquired during 2003–2005 (Weddell Sea region) and to the Radarsat-1 Antarctic Mapping Project (RAMP) mosaic, covering the Antarctic wide near-coastal zone. These images cover scenes under heterogenous backscattering signatures for all seasons with variable meteorological, oceanographic and acquisition parameters (e.g. band, polarization). Our method is highly adaptable to distinguish icebergs from ambiguous objects hidden in the images. The average false positive rate and miss rate are 2.3 ± 0.4% and 3.3 ± 0.4%, respectively. Overall, 9512 icebergs with sizes varying from 0.1 to 4567.82 km2 are detected with average classification accuracy of 97.5 ± 0.6%. The results confirm that the method presented here is robust for widespread iceberg detection in the Antarctic seas.