Seasonal Arctic sea ice forecasting with probabilistic deep learning

Arctic amplification has driven rapid declines in sea ice extent, with September extent now about half of 1979 levels and projections indicating potential ice-free summers by mid-century. This loss has profound ecological, societal, and climate impacts and may influence mid-latitude weather. Seasonal sea ice forecasting remains challenging: operational dynamical models often fail to beat simple statistical methods beyond 1–2 months lead time, despite studies indicating higher potential predictability. The research question is whether a data-driven, probabilistic deep learning approach can extend the range and reliability of seasonal Arctic sea ice forecasts and provide calibrated uncertainty useful for decision-making. The study introduces IceNet, trained on long climate simulations and observations to forecast six months ahead, aiming to outperform state-of-the-art dynamical models, especially for summer and extreme events, and to deliver well-calibrated probabilities enabling practical tools such as probabilistic ice edge bounds.

Prior work highlights limitations of seasonal forecasts from deterministic coupled atmosphere–ice–ocean models at lead times beyond ~2 months, due to chaotic atmospheric variability and processes like melt-season thickness that impose predictability barriers. Nevertheless, potential predictability is higher than realized in current systems. Deep learning has achieved success across Earth science applications, including remote sensing, and earlier sea ice studies used neural networks and convolutional approaches but with limited receptive fields and deterministic outputs. Standard practice defines the ice edge at 15% SIC. Dynamical systems like ECMWF SEAS5 are strong benchmarks but often require post hoc calibration. The Sea Ice Outlook (SIO) multi-model ensemble provides community baselines for September SIE. There is growing interest in probabilistic, calibrated forecasts that quantify uncertainty and in tools that bound classification frontiers for practical decision support.

Model: IceNet is an ensemble of 25 U-Net convolutional neural networks producing pixel-wise probabilistic forecasts for three SIC classes: open water (≤15%), marginal (15–80%), and full ice (≥80%) for each of the next six months on a 25 km EASE2 grid. Inputs are monthly-averaged fields comprising SIC, 11 climate variables (temperatures, radiation, winds, geopotential heights including tropospheric and stratospheric indicators), statistical SIC forecasts (a linear trend extrapolation), and metadata, stacked into 50 channels.

Architecture and training: Each U-Net follows an encoder–decoder with skip connections and batch normalization, totaling ~44M parameters. Training uses focal loss for class imbalance, Adam optimizer, He initialization, and month-wise loss weighting based on an active grid cell region that expands in winter and shrinks in summer to avoid dominance by trivial open-water cells. Outputs are logits per class per lead time, with post-hoc temperature scaling to calibrate probabilities. Ensemble-mean probabilities are computed by averaging ensemble member categorical distributions.

Data and preprocessing: Observational SIC from OSI-SAF (OSI-450/OSI-430-b) 1979–present are regridded to EASE2; polar hole gaps are bilinearly interpolated for inputs but excluded from targets. ERA5 monthly reanalysis provides non-SIC variables on surface and pressure levels; anomaly fields are formed by subtracting monthly climatologies (1979–2011), and all inputs are standardized using training-period statistics. SIC is scaled to [0,1]. Missing observational months (e.g., due to sensor gaps) are masked or removed as needed.

Transfer learning: To address limited observations, each ensemble member is pre-trained on 2220 years of CMIP6 data (MRI-ESM2.0 and EC-Earth3 ensembles; historical+SSP2-4.5, 1850–2100), shuffled across models and time to avoid overfitting to a single model’s physics. Validation during pre-training is performed on observations (2012–2017) to select checkpoints. Models are then fine-tuned on observations (1979–2011) with learning-rate scheduling, early stopping, and checkpointing with validation on 2012–2017. Test years (2018–2020) are held out.

Calibration: Temperature scaling is applied in two stages: first per ensemble member (single T across lead times), then on the ensemble mean (lead-time-specific T) by minimizing categorical cross-entropy on validation years via Brent–Dekker optimization. Ensembling and temperature scaling improve probabilistic calibration.

Benchmarks and evaluation: Benchmarks include ECMWF SEAS5 (25-member ensemble; bias-corrected using 2002–2011 mean error fields per month and lead time) and a grid-cell-wise linear trend SIC extrapolation. Model outputs are assessed via binary classification of SIC>15% using sea ice probability p=P(SIC>15%) with threshold 0.5; SEAS5 and trend SIC fields are similarly thresholded at 15%. Binary accuracy is computed over an active grid cell region and is related to integrated ice edge error by a normalization. Additional analyses include calibration curves (observed ice frequency vs predicted SIP), ice-edge bounding using calibrated probability contours p′ and 1−p′ to define probabilistic bounds, and permute-and-predict variable importance to quantify sensitivity to each input across months and lead times.

Implementation: TensorFlow in Python; Nvidia Quadro P4000 GPU; custom data loader for on-the-fly batching; Bayesian hyperparameter tuning (Weights & Biases) for initial learning rate, filters, and batch size. Inference speed is orders of magnitude faster than dynamical models once trained.

• Performance vs benchmarks: Over validation and test years (2012–2020), IceNet exceeds SEAS5 and linear trend binary accuracy at lead times of 2 months and beyond, with the largest gains in late summer to autumn (August–October). Heatmaps indicate improvements over SEAS5 up to about +2.9 percentage points in certain months/lead times and consistent gains over the linear trend, demonstrating skill beyond linear decline.
• Extreme events: For anomalous September SIE years (2012 lowest, 2013 high, 2020 second-lowest), IceNet’s September forecasts substantially outperform SEAS5 and linear trend at 4–2 month lead times. Example forecasts achieve high binary accuracies (e.g., 2012: 90.4% at 4 months, rising to 95.8% at 1 month; 2020: 91.5% at 4 months to 94.0% at 1 month) with relatively low SIE errors, showing strong capability for extremes.
• Seasonal dependence and predictability barrier: Skill dips for long-lead summer forecasts reflect the spring predictability barrier (importance of melt-season thickness), yet IceNet is competitive with or better than benchmarks during this period at 2–4 month leads.
• Calibration: IceNet’s SIP is near-perfectly calibrated over 2018–2020; observed ice frequencies align closely with predicted probabilities across bins. SEAS5 overestimates ice probability (notably many errors at predicted p=1), underscoring IceNet’s probabilistic reliability.
• Ice-edge bounding: Using calibrated probabilities, IceNet can probabilistically bound the observed ice edge between SIP contours p′ and 1−p′. Choosing p′≈0.036 bounds ~90% of the ice edge while covering ~24.4% of the domain across validation years/lead times. Outside this region, binary accuracy exceeds 99%, enabling a three-region segmentation (confident open water, ice edge region, confident ice) valuable for operational planning.
• Pre-training and ensembling: CMIP6 pre-training yields modest average gains (+0.26% binary accuracy across 2012–2020) with mixed effects by season (slightly negative for some long-lead September forecasts), suggesting benefits depend on climate model fidelity. Ensembling consistently improves accuracy (notably +0.6–1.4 percentage points in summer long-lead forecasts) and calibration; combined with pre-training, gains reach ~+1–2.4 percentage points depending on month and lead time.
• Variable importance: At short leads (1 month), initial SIC fields are most influential for both March and September targets. For September at 3 months, June initial conditions (including sea level pressure and 500 hPa geopotential height anomalies) are important, reflecting synoptic controls and mid-melt-season memory. At longer leads, dependence shifts toward linear trend inputs, consistent with diminishing initial-condition memory.
• Computational efficiency: Once trained, IceNet runs over 2000× faster on a laptop GPU than SEAS5 on a supercomputer, producing six-month probabilistic maps in under 10 seconds, enabling practical, rapid updates.

The study demonstrates that a probabilistic deep learning system can extend accurate seasonal Arctic sea ice forecasts beyond what leading dynamical models typically achieve, particularly at 2–4 month lead times for late summer and autumn and during extreme events. By directly producing calibrated probabilities, IceNet not only improves categorical accuracy but also quantifies uncertainty, enabling operational tools such as probabilistic ice-edge bounds that capture most of the true edge while maintaining high confidence elsewhere. These findings address the central research question by showing that data-driven models trained on large climate simulations and observations can learn physically plausible, seasonally varying relationships that complement and in some regimes surpass dynamical systems. The variable-importance patterns mirror known causal mechanisms and predictability timescales, reinforcing the physical credibility of the learned mappings. The performance gap with dynamical models in certain seasons/regions highlights where improved physics, data assimilation, and calibration in dynamical systems could yield the greatest benefits. The calibrated, sharp probabilistic outputs are directly relevant to risk-aware decision-making in shipping, ecosystem management, and potential downstream improvements in mid-latitude weather forecasts contingent on Arctic teleconnections.

IceNet, a deep learning ensemble trained via transfer learning and calibrated post hoc, delivers state-of-the-art seasonal Arctic sea ice forecasts that outperform a leading dynamical model (SEAS5) at lead times of two months and beyond, with strong skill for extreme September extents. Its probabilistic, well-calibrated outputs support a practical framework to bound the ice edge and define confidence regions, offering immediate utility for Arctic operations and conservation planning. The approach is computationally efficient, enabling rapid updates. Future work will assess adding sea ice thickness to improve summer forecasts, develop a daily, online version to enhance short-lead performance, and further integrate additional observations (e.g., melt ponds, ocean currents, waves) where available. Insights from IceNet’s variable importance can guide observational priorities and improvements in dynamical model physics and calibration.

• Predictability barrier: Skill decreases for long-lead summer forecasts due to the spring predictability barrier (importance of melt-season thickness), limiting precision of probabilistic ice-edge bounds.
• Input limitations: Not all relevant processes are observed consistently over 1979–2020 (e.g., waves, ocean currents, melt ponds), so these were excluded and may limit forecast skill.
• Monthly inputs: Using monthly-averaged inputs reduces sensitivity to weather-scale phenomena and likely contributes to IceNet underperforming SEAS5 at 1-month lead.
• Transfer learning dependence: CMIP6 pre-training offers modest average gains and can be detrimental for some months/leads, reflecting limitations in climate model representations (e.g., summer melt processes).
• Data issues: Satellite SIC retrieval uncertainties (especially near coastlines and during summer) introduce noise; the 41-year observational record constrains purely data-driven learning; polar hole interpolation is used for inputs but excluded from targets.
• Calibration and sharpness trade-offs: Bounding precision depends on forecast sharpness; reliability of bounds is maintained partly by inflating the edge region as lead time increases, limiting spatial tightness when inherent predictability is low.