A Deep Learning Approach for Lake Ice Cover Forecasting

Abstract. Lakes cover a significant proportion of the high-latitude landscape and exert a strong influence on local weather and climate. Their seasonal lake ice cover (LIC) further impacts lake-atmosphere interactions, while also providing key socioeconomic services for northern communities. Climate change is impacting LIC and its thickness, two thematic products of Lakes as an Essential Climate Variable (ECV). Accurate prediction of LIC improves numerical weather prediction (e.g. lake-effect snowfall and thermal moderation) and is crucial for anticipating the impacts of climate change in lake-rich regions of the Northern Hemisphere.

This paper introduces LIF-DL (Lake Ice Forecasting using Deep Learning), a novel data-driven model for forecasting LIC extent across entire lake surfaces. LIF-DL uses Spatial-Temporal Transformer Networks (STTN) to capture relationships between lake conditions (ice and open water), lake depth and atmospheric forcings. The study focuses on five large Canadian lakes with pronounced ice phenology: Great Slave Lake, Great Bear Lake, Lake Winnipeg, Lake Athabasca, and Reindeer Lake. Data sources included ice cover observations from the Interactive Multi-Sensor Snow and Ice Monitoring System (IMS), atmospheric reanalysis from the European Centre for Medium-Range Weather Forecasts (ECMWF) 5th generation of European ReAnalysis (ERA5 and ERA5-Land), and Canadian Ice Service (CIS) records for external validation. To benchmark the proposed approach against a traditional physics-based model, the widely used Freshwater Lake (FLake) model embedded in ERA5 and ERA5-Land was employed. LIF-DL was trained to produce one-week forecasts using data from 2004–2017 and then deployed auto-regressively to predict ice cover during the 2018–2022 holdout period. Forecasts were evaluated against IMS and CIS observations and compared with those from FLake.

Across all evaluations—phenology timing, ice cover fraction, and spatial patterning—LIF-DL consistently outperformed FLake. Freeze-up and break-up events were predicted within 3–9 days of observations (versus 5–22 days for FLake), and ice cover fraction (range 0–1) root mean squared errors were reduced (0.06–0.16 versus 0.1–0.2). A key advantage of LIF-DL was its capacity to represent spatial dependencies across lake surfaces, producing coherent freeze-up and break-up dynamics and realistic spatial clustering of early and late ice timing compared to the fragmented patterns of FLake. These improvements reduced extreme timing biases—from as much as 30 days to only 4–6 days—particularly for large, deep lakes. Variable importance analysis indicated sensitivity to physically meaningful drivers, including air temperature, accumulated degree days, solar radiation, and lake depth, suggesting that LIF-DL learned relevant physical processes rather than statistical artifacts. Finally, the model maintained stable performance when iteratively forecasting over a four-year period, demonstrating robustness under varying atmospheric conditions.

The demonstrated accuracy, robustness, and physical interpretability of LIF-DL highlight the potential of deep learning for advancing lake ice modelling. Future research should focus on integrating physical constraints to develop hybrid physics-machine learning frameworks, improving model interpretability, and expanding to new predictive variables such as ice thickness and snow cover. Leveraging emerging high-resolution satellite datasets will further enhance spatial fidelity and enable application to smaller lakes. Ultimately, spatiotemporal deep learning represents a transformative step toward next-generation, spatially resolved lake ice forecasts that can improve weather and climate prediction, inform northern transportation planning, and support climate change adaptation in lake-rich regions of the Northern Hemisphere.