Using machine learning approaches to forecast seasonal sea ice conditions in Hudson Bay


icon for download PDF here Andrea Scott K. Andrea Scott, Department of Systems Design Engineering

ZG Zacharie Gousseau, Department of Systems Design Engineering


Accurate and timely forecasts of sea ice conditions are critical for safe shipping in the Canadian Arctic. While short-term forecasts are important for day-to-day operations, longer-term seasonal forecasts are useful for strategic planning by shipping companies and offshore operators. These seasonal forecasts are becoming increasingly important given declining sea ice cover in the Arctic over recent decades.

Seasonal forecasting approaches have historically used statistical or dynamical approaches. More recently, convolutional neural networks, which are able to learn nonlinear relationships between spatial patterns in input data, have been used for sea ice prediction but have been unable to produce long-term forecasts that propagate in space and time. This study applies a novel spatiotemporal forecasting method based on a sequence-to-sequence machine learning approach to provide daily forecasts of the probability of sea ice presence in Hudson Bay, Canada with lead times up to 90 days. In addition, the study tests the ability to predict freeze-up/breakup dates within a seven-day period at specific locations of interest to shipping operators and communities.


The study focused on the Hudson Bay, James Bay, Hudson Strait, and Foxe Basin system, a region bordered by 39 communities, 29 of which are only accessible by sea or air. The area is seasonally covered by first-year ice, with open water over most of the area each summer. Shipping traffic, mostly confined to ice-free and shoulder seasons, is generated by mining, fishing, tourism, sea-lift and research activities (Figure 1).

The study region with locations of interest

Figure 1: The study region with locations of interest. Insets show locations of ports (red), the nearest model grid points (blue) and a bounding box (blue) that approximates a grid cell.

The study used ERA5 reanalysis data produced by the European Centre for Medium-Range Weather Forecasting (ECMWF) for model predictors and validation and included eight input variables over the period of 1985 to 2017: sea ice concentration, sea surface temperature, 2 m air temperature, surface sensible heat flux, wind 10 m U component, wind 10 m V component, land mask and additive degree days. Regional ice charts from the Canadian Ice Service were used for verification of freeze-up and breakup dates.

The forecast model architecture was conceptualized as a spatiotemporal sequence solvable by a general sequence-to-sequence machine learning framework where the encoder component transformed inputs to an encoded state of fixed shape and the decoder component used the encoded state to generate sea ice presence probability within a 90-day forecast period. Two prediction models were developed: a “basic model” using geophysical data as input and an “augmented model” using corresponding Climate Normal data as additional inputs. A custom recurrent neural network decoder extrapolated states across time steps (Figure 2).

network architecture

Figure 2: (a) Overall network architecture and (b) custom decoder. In panel (a), the black portion refers to the Basic model, the red portion to the Augmented model. The dashed arrows show a process carried out only once (the initialization of the adder). FPN: feature pyramid network, ConvLSTM: convolutional long short-term memory network, NiN: network in network module, ReLU: rectified linear unit.

A training and validation protocol was developed to provide users with an operational tool that assessed forecasting skill. For each month a separate model was trained on data from the given month in addition to the preceding and following months. The model was trained over various time periods (years, months) each generating model weights for future training periods representing learned parameters that transformed the input to the output. The protocol was used to produce forecasts of sea ice presence for years 1996 to 2017 (Figure 3). The binary accuracy, Brier score, and accuracy of freeze-up and breakup dates were used to evaluate the performance of the machine learning predications against ERA5 sea ice concentration observations and Climate Normal baselines defined as the average of the ERA5 sea ice presence from 1985 to the last year in the training set for each experiment.

testing protocol

Figure 3: Training, validation and testing protocol for Basic and Augmented models.


The spatiotemporal sea ice forecasting method tested was capable of predicting sea ice presence probabilities with skill during the sea ice break up period in spring (May to July) in comparison to the Climate Normal forecasts. Results during the freeze-up period in fall (October to December) were mixed, with higher accuracy at short lead times demonstrated in November as compared to Climate Normal, but with decreasing accuracy at longer lead times (Figures 4, 5 and 6).

Binary accuracies

Figure 4: Binary accuracies as a function of lead time. Top row panels a–c show the binary accuracy of each model, while bottom row panels d–f show the differences in binary accuracy between the models.

Freeze up 

Figure 5: Accuracy of the predicted freeze-up date within 7 days.

Binary accuracy as a function of lead day for forecasts 

Figure 6:  Binary accuracy as a function of lead day for forecasts launched in (a) May, (b) June, (c) October, and (d) November.

Figure 7 maps the spatial probability of ice for three dates during the breakup period. The predicated pattern of breakup was found to be in good agreement with observations, in particular for the Augmented model.

Spatial patterns of sea ice

Figure 7: Spatial patterns of sea ice presence during breakup

The assessment of freeze-up and breakup date forecasting performance against EAR5 baseline observations show similar spatial patterns for the Basic and Augmented models at 60 lead days, with differences between the two models at 30 lead days. The Climate Normal freeze-up accuracy for 30 lead days was significantly different than other models (Figure 8).

Freeze up

Figure 8: Accuracy of the predicted breakup date within 7 days.


The study applied a novel deep learning approach to seasonal sea ice forecasting in Hudson Bay, Canada. A spatiotemporal forecasting method based on sequence-to-sequence learning provided daily spatial maps of sea ice presence probability with lead times up to 90 days and freeze-up/breakup dates of strategic use to shipping operators. The proposed method was capable of predicting sea ice presence probabilities with skill during the breakup season in comparison to both Climate Normal and sea ice concentration forecasts from a leading subseasonal-to-seasonal forecasting system. The model was demonstrated in hindcast mode to allow for the evaluation of forecasted predications; however, the design allows the approach to be used as a forecasting tool. Compared to dynamical forecasting systems, the proposed approach is very time efficient as once the initial model is trained the fine-tuning process for new inputs is minimal. 

A study limitation is that it relies on data from reanalyses such that, without an additional downscaling module, the spatial resolution of forecasts cannot exceed that of the input data, which was 31 km. It was noted, however, that this resolution was similar to that used in other studies on seasonal forecasting for marine transportation. Future work includes an expansion of experiments over the entire Arctic region, beginning in the Beaufort Sea, deploying ensemble methods using more recent deep learning architectures.

Asadi, N., Lamontagne, P., King, M., Richard, M., and Scott, K. A. (2022). Probabilistic spatiotemporal seasonal sea ice presence forecasting using sequence-to-sequence learning and ERA5 data in the Hudson Bay region. The Cryosphere, 16, 9.

For more information about WaterResearch, contact Julie Grant.

Sea ice photo credit: Andrea Scott