Predicting the timing of annual river ice breakup is crucial for residents to prepare for potential flooding and assess the safety of rivers for transportation. This analysis develops a deep learning approach using meteorological and geospatial data products to forecast river ice breakup. We selected 33 locations along eight major rivers across Alaska, USA, and Western Canada, leveraging annual breakup dates from the Alaska-Pacific River Forecast Center database. Daily meteorological data from Daymet, along with static watershed attributes from the pan-Arctic catchment database, were used to develop a Long Short-Term Memory model (LSTM) for predicting river ice breakup. Of the 33 locations, 23 were used for training the LSTM. The model demonstrated high efficacy, accurately predicting the annual breakup date with a mean absolute error (MAE) of 5.40 days, a standard deviation of 4.03 days and a mean absolute percentage error (MAPE) of 4.37%. The spatial generalizability of the LSTM was evaluated using the remaining 10 locations as holdouts, with eight of the 10 locations averaging a MAPE of less than 8% over the entire time series (1980 to 2023). Additionally, we retrieved 51 long-range seasonal forecast ensembles from the Copernicus Climate Data Store and applied trained LSTM to them to showcase the capability of the LSTM to predict future river ice breakup using operational weather forecasts. To analyze marginal contribution of LSTM inputs for predictions, Shapley values were calculated. A new temporal correction scheme was developed and applied to Shapley values to address unique features of the breakup data.
