Modeling daily ice cover in Northern Hemisphere lakes with Long-Short Term Memory Networks
Abstract
Over recent decades, various models have been developed to quantify global lake ice loss in the past and future. However, these models cannot resolve the daily variability of ice cover at appropriate spatial scales. Here, we present an application of a Long-Short Term Memory (LSTM) model, trained with Landsat observations during 1984-2012, to predict daily ice cover changes for lakes in the Northern Hemisphere. The LSTM model achieved a mean Nash-Sutcliffe Efficiency of 0.78 (median of 0.94) during the test period (2013-2022), indicating a strong capability of extrapolating predictions over time in most regions. The model derived ice phenology, including ice-off,ice duration in-situ observations matched well against mean absolute errors of 11.6, 6.3 and 14.3 days respectively. Given the high accuracy in re-creating historical conditions, the model can be potentially used for future projections to enhance our understanding of global lake ice loss under climate change.
