TPFMNet: A Probabilistic U-Net With Triple-Path Fusion Module for Spatially Adaptive Uncertainty in Year-Round Arctic Sea Ice Forecasting

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:A novel deep learning method for spatiotemporal forecasting with uncertainty quantification that advances physical climate modeling.

Accurate year-round Arctic sea ice concentration (SIC) forecasting with reliable uncertainty quantification is critical for polar navigation and climate risk assessment. While deep learning (DL) has significantly improved predictive accuracy, existing deterministic models often fail to account for the extreme nonstationarity and inherent noise in satellite-derived sea ice products, particularly during extreme seasonal transitions such as the spring predictability barrier (SPB). In this study, we propose TPFMNet, an efficient probabilistic U-Net-based framework designed for spatially adaptive uncertainty estimation in year-round Arctic SIC forecasting. The core of the architecture is the tri-path fusion module (TPFM), which explicitly decouples multiscale physical dynamics into three synergistic pathways: 1) dual-scale neighborhood attention for capturing highly localized kinematic fractures; 2) zero-inference-cost structurally reparameterized convolutions for resolving broad morphological patterns; and 3) a global frequency-domain mixer for tracking basin-scale seasonal periodicities. In addition, a geographically guided coordinate attention mechanism embeds spatial priors to accurately monitor the dynamically sensitive marginal ice zone (MIZ). To quantify heteroscedastic aleatoric uncertainty without the computational burden of ensemble methods, TPFMNet models the predictive distribution as a Laplace distribution, optimized via negative log-likelihood (NLL). Extensive experiments on OSI SAF historical datasets demonstrate that TPFMNet achieves state-of-the-art predictive performance, significantly outperforming recent spatiotemporal and domain-specific baseline models. Specifically, it yields a superior overall mean absolute error (MAE) of 0.01670 and an $R^{2}$ of 0.9603. The probabilistic outputs exhibit exceptional calibration, maintaining a prediction interval coverage probability of over 95% while generating sharp, informative bounds. Notably, the estimated uncertainty is physically consistent with Arctic cryospheric processes, adaptively intensifying during the highly volatile summer melt season and explicitly mitigating the SPB. These results establish TPFMNet as a robust, highly efficient tool for operational Arctic maritime applications. The code will be made publicly available at: https://github.com/wmingqiang01/TPFMNet

• semanticscholar https://doi.org/10.1109/tgrs.2026.3708289first seen 2026-07-17 05:32:16