Estimating Floating Ice Coverage in Tidewater Glacier Bays Automatically from Aerial Imagery

This study focuses on the estimation of floating ice distribution on the sea surface from aerial imagery at tidewater glaciers. The presence of ice-mélange in the glacial bay affects ice-loss at the glacier terminus via buttressing of the glacier face, insulation from ocean heat and breaking of oceanic currents. Thus, estimation of its formation, evolution, distribution and dissipation will aid an understanding of climate-change mechanisms at glacial bays. Given the challenges stemming from lack of data and labeled training samples, we employ a pre-trained vision model with zero-shot performance, the Segment Anything Model (SAM). SAM is fine-tuned to perform automatic estimation of floating-ice distribution on a dataset of images collected in Svalbard at various locations and through various modalities, by segmenting the images into floating ice and ocean surface not covered by ice. The Low-rank adaptation technique is employed on the image encoder structure to reduce the training data and hardware resources needed for model fine-tuning. This approach opens new avenues for real-time analysis of glacial dynamics and their implications on global climate patterns.