Detecting Sea Ice Thickness from Close-Range Images using Machine Learning
Alessandro Toffoli
This study introduces a novel image-based approach for measuring sea ice thickness using close-range imagery collected during Voyage 1 of the 2023–24 Antarctic season, focusing specifically on overturning ice events caused by icebreaker navigation. These events briefly expose ice edge profiles, offering valuable thickness information. A custom pipeline leveraging the Segment Anything Model (SAM) from Meta AI was fine-tuned to automatically detect overturning ice and extract thickness data from orthorectified images, significantly reducing manual observation efforts. Addressing limitations of satellite-based thickness measurements and labor-intensive conventional in-situ observations (ASPeCt protocol), this approach facilitates scalable and cost-effective sea ice monitoring. The SAM-based model demonstrated strong segmentation accuracy, closely aligning with expert ASPeCt observations. This marks the first application of fine-tuned SAM for automated overturning ice detection in shipborne images, establishing a broadly adaptable framework for future environmental monitoring and remote sensing tasks.
Please use this thread for discussion about this talk.