Sea ice detection from concurrent visible and SAR imagery using a convolutional neural network

REGISTER NOW

Abstract:

Accurate measurements of Antarctic sea ice extent are important for safe navigation of polar ships, understanding ecosystem dynamics and identifying sea ice response to changing temperature and weather patterns. Sea ice has primarily been detected using passive microwave satellite data; however, the coarse spatial resolution of this imagery precludes its use in decision-making tasks. Recent advances in the spatial coverage and temporal resolution of Sentinel-1 Synthetic Aperture Radar (SAR) and multispectral visible imagery (MODIS) are opening new opportunities to rapidly detect sea ice extent with greater positional precision. Both image types exhibit strengths and limitations when used in isolation to detect sea ice extent. The MODIS platform provides daily imagery of the whole Antarctic continent, but cloud and sunlight conditions can inhibit the detection of sea ice. SAR can distinguish ice from water independent of cloud and lighting conditions, but the imagery contains speckle and ambiguous signals caused by ocean waves and ice surface meltwater.

This presentation will describe the development of a novel supervised convolutional neural network (CNN), CombiNet, trained to detect sea ice extent in concurrent MODIS and SAR imagery. The performance of CombiNet is compared to two separate CNN trained on MODIS and SAR imagery individually. This presentation will discuss how the fusing of these two image types can address the abovementioned limitations for sea ice detection alongside considerations when training a CNN on multiple concurrent satellite data sources.

Bio:

Martin is a machine learning researcher in the Artificial Intelligence (AI) Lab at the British Antarctic Survey (BAS). His research primarily focusses on the application of machine learning techniques to detect features in satellite imagery, including multispectral visible and Synthetic Aperture Radar (SAR) datasets. Martin has previously trained and applied AI techniques, including convolutional neural networks, to automatically detect coastal features and land covers in satellite imagery. His current research focusses on the detection and classification of sea ice and other polar features in SAR imagery. Martin is also the project manager for the AI4EO Guided Team Challenges, which forms part of the taught element of the Application of AI to the study of Environmental Risks (AI4ER) CDT.