Sea ice floe segmentation in close-range optical imagery using active contour and foundation models

TL;DR

This work addresses the challenge of autonomously segmenting sea-ice floes in close-range optical imagery from the Antarctic marginal ice zone, where boundaries are often blurred by heterogeneity and low contrast. It systematically compares a traditional gradient vector flow (GVF) active contour method, a vision foundation model (Segment Anything Model, SAM) in automatic and prompt-based modes, and a hybrid SAM-GVF approach on a large, high-resolution dataset collected from an icebreaker, benchmarked against extensive manual segmentation. The results show SAM-P as the most robust and efficient option for real-time or large-scale analyses, while SAM-GVF offers superior boundary accuracy at a substantial computational cost; GVF generally underperforms in dense or diffuse boundary conditions. The study provides a reproducible, automated pipeline and a publicly available benchmark to enable scalable quantitative analyses of floe size distribution and ice concentration, with implications for improving MIZ modeling and monitoring using field imagery.

Abstract

The size of sea ice floes in the marginal ice zone (MIZ) is a key factor influencing ice coverage, albedo, wave propagation, and ocean--atmosphere energy exchanges. Floe size can be observed by processing visual-range imagery from ships, aircraft, or satellites. However, autonomously capturing floe boundaries remains challenging, particularly due to sea ice heterogeneity, which impairs boundary definition and reduces image clarity. This study evaluates the accuracy of sea ice floe segmentation using the gradient vector flow (GVF) active contour method, the deep learning-based Segment Anything Model (SAM), and a hybrid approach combining GVF and SAM. Methods are evaluated on a representative subset of a large dataset of close-range, high-resolution imagery collected from cameras aboard an icebreaker during an Antarctic winter expedition. Spanning a wide range of ice conditions and image clarity in the MIZ, the subset provides a rigorous segmentation test bed. Performance is assessed in terms of floe detection accuracy, size distribution, and ice concentration, with results compared against a manually segmented benchmark. Results indicate SAM, in prompt-driven mode, offers the best balance between accuracy and computational efficiency. Its strong performance in estimating sea ice concentration and detecting floes, while maintaining close agreement with benchmark floe size distributions, makes it suitable for real-time applications and scalable analyses of large imagery datasets. Compared with SAM, the combined SAM-GVF method provides more accurate floe boundary delineation, although at much higher computational cost, and is therefore better suited for analyses requiring precise floe shapes.

Figures (13)

• Figure 1: Antarctic SCALE expedition, winter 2022: (a) camera mounted on S.A. Agulhas II; (b) ship track from departure on 11 July 2022 to return on 31 July 2022, entering the sea ice region on 19 July and exiting on 24 July, overlaid on sea ice concentration from the University of Bremen AMSR2 ASI product spreen2008sea; (c--e) high-resolution images of the sea ice surface captured under (c) conditions featuring distinct ice floes with defined edges, (d) diffuse sea ice boundaries characterized by densely packed, irregularly shaped floes with indistinct or visually merged edges, and (e) at nighttime.
• Figure 2: Sample sea ice image (a) from the Antarctic SCALE expedition, winter 2022, after perspective correction, highlighting the region of interest for analysis in red (b). The selected region after image enhancement, including gray-level homogenization through Gaussian bilateral filtering, anisotropic diffusion, and contrast enhancement using CLAHE (c).
• Figure 3: Uncertainty analysis of extrinsic camera parameters on average floe diameter, with the floes used in the analysis highlighted in red. Variations include (a) $\pm$ 1 m elevation, (b) $\pm$ 1$^{\circ}$ tilt, and (c) $\pm$ 2$^{\circ}$ roll. Floe diameters are normalized against values derived from orthorectification using ship-measured extrinsics. The shaded blue area indicates a 5% uncertainty range.
• Figure 4: Illustration of the GVF Snake algorithm workflow: (a) enhanced sea ice image; (b) seeds (red dots) representing floe centers on the binarized image; (c) floe perimeters (red lines) identified from isolated, single-seed objects; (d) elliptical initial contours (red lines) of floes containing a single seed after morphological erosion; (e) circular initial contours (red lines) for remaining unused seed; (f) floes identified directly from the binarization process (as shown in Fig. \ref{['fig:gvf_outputs']}c); (g) floes detected by the GVF snake algorithm initialized from elliptical contours; (h) floes detected by GVF initialized from circular contours; (i) final binary segmentation with all detected floes.
• Figure 5: SAM segmentation modes applied to enhanced sea ice imagery. (a--c) Automatic segmentation mode (SAM-A): (a) grid-based fixed points automatically selected by SAM; (b) consolidation of masks returned by the model after filtering; (c) final unified binary mask of detected floes. (d--f) Prompt-based segmentation mode (SAM-P): (d) seed points (red dots) marking floe locations; (e) consolidation of masks returned by the model after filtering; (f) final unified binary mask of detected floes.