researchsquare.comSource publication:
Loading content…
Arctic sea ice concentration forecasting has been limited by the difficulty of fusing multi-timescale information, with statistical models missing nonlinear relationships and numerical models depending on initial conditions. STICEnet replaces the U-Net encoder-decoder path with strictly unidirectional temporal causal and transposed convolutions and embeds channel and spatial attention in the skip connections, trained on NSIDC sea ice concentration and ERA5 atmospheric and oceanic variables. Across one- to six-month leads its RMSE, MAE and R2 all beat CNN, ConvLSTM, IceNet and MT-IceNet, reaching 7.51% RMSE and 0.925 R2 at month six.
One- to six-month Arctic sea ice concentration forecasts now have a model that outperforms the listed baselines at every lead time, with error no longer degrading sharply as the horizon grows. Earlier deep models accumulated error at long leads, with ConvLSTM and multi-task ConvLSTM reaching 11.07% and 15.79% RMSE and IceNet dropping to 0.403-0.567 R2. On a dataset built from NSIDC sea ice concentration and ERA5 reanalysis variables, STICEnet reaches 5.36% RMSE and 1.236% MAE at month one and 7.51% RMSE, 1.928% MAE and 0.925 R2 at month six, below MT-IceNet's 5.50%, 1.313% and 7.77%, 1.967%, 0.915.
The model treats the time direction as a physical constraint, letting the encoder see only past time steps and restoring the temporal scale in the decoder. Ordinary convolution reads both directions along time and mixes future steps into current features, inflating predictions during fast transitions such as the spring melt. The temporal causal convolutional encoder uses causal convolution with gating and residual connections to extract features that depend only on historical steps, while the temporal transposed convolution decoder restores temporal resolution with transposed convolution, batch normalization and pooling; ablation shows adding this mechanism alone lowers RMSE at one, three and six months from 5.50%, 7.61% and 7.77% to 5.39%, 7.48% and 7.53%.
Local abrupt changes at the ice edge and long-range dependencies of the ice pack are handled inside one multi-scale encoding-decoding structure. Earlier methods struggled to fuse multi-timescale data jointly, often handling local ice-edge change and global ice-sheet trends in separate modules. The encode block first extracts spatial features with multi-level convolution and pooling, then models dynamic temporal dependencies with a state-space model, and the decode block upsamples and progressively restores detail; ablation shows R2 rising to 0.966, 0.926 and 0.923 at the three horizons.
Weights on atmospheric and oceanic driver variables are calibrated adaptively by attention over channels and spatial locations rather than fixed. The contribution of multi-source drivers varies by region and season, so fixed weights leave persistent bias in marginal ice zones. Channel attention reweights multi-source feature channels and spatial attention focuses on key regions such as the ice edge; ablation shows MAE falling to 1.257% at month one and 1.897% at month three.
The model is trained and evaluated on a 1979-2021 dataset built from NSIDC sea ice concentration and ERA5 reanalysis variables, covering one- to six-month lead times and outputting pixel-level sea ice concentration on a 25 km polar stereographic grid. Training uses 408 months from 1979 to 2015 and testing uses 96 months from 2016 to 2018, with an additional check on the unseen January 2019 to December 2021 period. The intended readers are researchers and engineers working on polar sea ice forecasting and spatiotemporal sequence modeling.
The model captures interannual sea ice decay trends and anomalous responses to extreme climate forcing inadequately, which the authors attribute to limited modeling capacity for slow variables and nonlinear cumulative effects in the climate system. During the 2016-2017 super El Nino period, long-lead forecasts showed localized positive deviations; during the 2019-2021 summer minima, predicted September sea ice extent leveled off around 6.2-6.5 million square kilometers against observed values of about 4.5 million and 4.1 million square kilometers, which the authors attribute to Siberian Arctic heatwaves and positive ice-albedo feedback causing sub-grid localized melting. The authors propose building a coupled sea ice anomaly and meteorological anomaly prediction framework and incorporating more expressive time-series modules.