Answers are grounded in this paper and cite their sources.
Ground-based Global Navigation Satellite System interferometric reflectometry (GNSS-IR) provides frequent signal-to-noise ratio (SNR) observations for coastal sea ice monitoring. However, conventional physical indicators show substantial overlap between sea ice and open water at the individual-arc scale, while satellite imagery provides reference information at coarser temporal and spatial scales. We develop a daily mean multiple-instance learning (DM-MIL) framework that aggregates SNR waveform features to predict daily ice dominance while retaining information from individual arcs. The model was trained using observations from 2019 to 2020 at the TUKT station in Tuktoyaktuk, Canada, and independently evaluated on the 2021 dataset. It achieved 99.72% classification accuracy over 351 image-supported stable-state days and captured the seasonal progression of melt and freeze-up within image-defined transition periods. Arc-level ice-likeness scores showed 91.60% agreement with the binarized daily reference, while their daily averages achieved 99.18% agreement across all 365 days of 2021. Seasonal variations in conventional physical indicators were consistent with the model outputs, providing signal-level evidence that the predictions were associated with measurable changes in the reflected signals. By aligning model supervision and prediction with daily reference information, DM-MIL provides a reliable daily monitoring product and complementary sub-daily information. The framework highlights the potential of existing coastal GNSS networks to complement satellite remote sensing and extend continuous monitoring of nearshore sea ice conditions.
No takes yet. Share an insight, caveat, or question.
Yuan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: