Deep Learning for Seasonal Navigability Prediction Along the Northern Sea Route: When Does It Add Value?

The Northern Sea Route (NSR) is becoming increasingly accessible as Arctic sea ice declines, motivating data-driven forecasts of seasonal navigability. We compiled a 13-year (2013–2025) monthly dataset of AMSR2 sea ice concentration (SIC) and ERA5 atmospheric reanalysis variables over the NSR corridor (68–80° N, 30–180° E) and benchmarked a hierarchy of forecasting models for 1-, 3-, and 6-month lead times. Baselines (climatology, persistence, anomaly persistence, SARIMA, ridge regression) were compared with compact deep learning architectures (LSTM, Transformer; 10,000–70,000 parameters) trained on 12-month sequences with anomaly targets and five-seed ensembles. Three findings emerge. First, the seasonal cycle explains 98.0% of the monthly SIC variance, so climatology alone yields RMSE = 4.56% and three-class navigability accuracy of 87.5%. Second, SARIMA, the compact LSTM ensemble, random forest, and MLP_small all yield small positive skill scores over climatology: SARIMA achieves the lowest 1-month RMSE (3.98%, skill score +0.239), while the compact LSTM ensemble shows positive skill at all horizons (mean skill score +0.038); however, the bootstrap confidence intervals overlap and these differences are not statistically distinguishable from climatology. Third, all skilful models converge to identical classification metrics (accuracy 0.875, macro-F1 0.78, κ = 0.76); McNemar tests and overlapping bootstrap confidence intervals show no statistically significant differences. Permutation importance confirms that AMSR2 ice-state features dominate, whereas the high raw correlations of ERA5 radiation variables collapse after detrending. These results indicate that compact statistical and deep learning models are equivalent for NSR seasonal navigability prediction and that honest baseline comparison is essential when seasonal cycles dominate.