Using Machine Learning to Predict Complete Winter Ice Cover of a Freshwater Lake

Ice records at Lake George, an oligotrophic, dimictic freshwater lake in upstate New York, USA reveal that it has failed to freeze over completely 11 times since 1912, ten of which have occurred since 1990. This transition from annual to intermittent ice cover is analogous to many other dimictic freshwater lakes globally, however a unique characteristic for Lake George is that neither the onset time nor the duration of complete ice coverage has notably changed. Over sixty years of meteorological observations from a nearby airport are analyzed and a complicated picture emerges when considering the specific characteristics of each year (e.g., Lake George froze over in 1983 and 2007 despite the air temperature having a net warming effect on the lake in the month prior to ice-in). Such complexity makes the prediction of complete ice-coverage challenging, and we hereby present a machine learning approach to understand and predict the presence of complete ice-coverage on Lake George using only weather data. The machine learning classifiers performed adequately compared to observations, with one configuration having an overall accuracy of 91% and a root mean squared error of 10.5 days for ice-in and 5.5 days for ice-out. 21st Century predictions using data from a coupled-climate model (GFDL-CM3) were attempted, however the climate model failed to simulate historical regional weather with sufficient accuracy to adequately train the classification models. The success of the shorter-term predictions demonstrates the utility of machine learning in understanding the dynamics of ice on lakes, while the failure of longer-term predictions demonstrates the challenges of connecting climate-scale dynamics to local phenomena.