Region-level labels in ice charts can produce pixel-level segmentation for Sea Ice types

TL;DR

Pixel-level sea ice type labeling is hindered by scarce high-resolution ground truth. The paper presents a weakly supervised segmentation pipeline that learns per-pixel ice types from region-based ice-chart labels by encoding four classes into regional distributions and training with a polygon-level loss. It demonstrates higher class-wise $R^2$ scores across open water, young ice, first-year ice, and multiyear ice on the AI4Arctic dataset, outperforming a fully supervised U-Net and the AutoIce top solution. This approach enables high-resolution, operational sea ice mapping without pixel-level ground truth and suggests paths for applying the method to additional datasets and architectures.

Abstract

Fully supervised deep learning approaches have demonstrated impressive accuracy in sea ice classification, but their dependence on high-resolution labels presents a significant challenge due to the difficulty of obtaining such data. In response, our weakly supervised learning method provides a compelling alternative by utilizing lower-resolution regional labels from expert-annotated ice charts. This approach achieves exceptional pixel-level classification performance by introducing regional loss representations during training to measure the disparity between predicted and ice chart-derived sea ice type distributions. Leveraging the AI4Arctic Sea Ice Challenge Dataset, our method outperforms the fully supervised U-Net benchmark, the top solution of the AutoIce challenge, in both mapping resolution and class-wise accuracy, marking a significant advancement in automated operational sea ice mapping.

Figures (4)

• Figure 1: The architecture of the proposed pixel-level ice type mapping model based on a U-Net and the weakly supervised learning scheme (indicated by dash lines). The digit below each layer indicates the number of filters for each convolution/deconvolution layer.
• Figure 2: Visualization of results obtained from two SAR scenes. (a): SAR image in HH channel, file ID name: 20180707T113313cis, collected on July $7^{th}$, 2018 in Hudson Bay (center latitude/longitude: $57.07^{\circ}N, 77.96^{\circ}W$); (b): The HV channel of the SAR image in (a); (c): Classification results of (a) from the benchmark U-Net model (dark blue: open water; light blue: young ice; yellow: first year ice; red: multiyear ice); (d): Classification results of (a) from the proposed weakly supervised U-Net; (e): SAR image in HH channel, file ID name: 20190926T152005cis, collected on September $26^{th}$, 2019 in Western Arctic (center latitude/longitude: $72.63^{\circ}N, 126.61^{\circ}W$); (f): The HV channel of the SAR image in (e); (g): Classification results of (e) from the benchmark U-Net model; (h): Classification results of (e) from the proposed weakly supervised U-Net. The land area is masked in white.
• Figure 3: (a) Overview of the ice egg code parameters in the simplified egg code produced from the original ice charts. Concentration values range from 0 (empty) to 10 (fully- covered) with each increase of 1 representing a 10 percent concentration increase. (b) An example Ice chart showing eggcode and conversion of eggcode to ground truth tensor.
• Figure 4: Zoom-in examples of regions cropped from SAR scenes and corresponding predictions in Fig. \ref{['result visual3']}. First column: a region located within a polygon in Fig. \ref{['result visual3']} dominated by first-year ice. Second column: a region located within a polygon in Fig. \ref{['result visual3']} consisting of both young ice and multiyear ice. First row: regions in HH channel; Second row: regions in HV channel; Third row: overlay of HH channel and predictions from benchmark model, with the color bar provided in Fig. \ref{['result visual3']}; Fourth row: overlay of HH channel and predictions from weakly supervised model.