Skip to main content
U.S. flag

An official website of the United States government

Forecasting River Ice Breakup in Alaska USA Using a Long Short Term Memory Model

Presentation Date
Wednesday, December 13, 2023 at 2:10pm - Wednesday, December 13, 2023 at 6:30pm
Location
MC - Poster Hall A-C - South
Authors

Author

Abstract

Accurately forecasting when annual river ice breakup will occur across arctic rivers is of enormous significance for the ecosystems and inhabitants of those regions. The annual breakup leads to ecological destruction, damage to property and potentially even deaths. Forecasting when the breakup and subsequent ice jam induced flooding will occur, can help mitigate the impacts of these events. We studied the ice breakup events in rivers across Arctic Alaska and developed a machine learning model to forecast the timing of river ice breakup using meteorology and watershed characteristics. We collected the river ice breakup records within the National Weather Service (NWS) database and identified the eight locations with the longest and most complete time series for recorded river ice breakup. We also collected meteorological data at the sites from the DAYMET dataset. A long short term memory model (LSTM) was used to produce pseudo likelihood functions showing which day of each year had the highest probability of being the true breakup event. Using maximum likelihood estimation on the LSTM outputs, we were able to predict the breakup date for river ice within an average of 5.6 days using DAYMET meteorological data as input. While river ice processes such as freezeup and breakup are not represented in the global Earth System Models, to understand the state of river ice breakup process in the future, we identified/derived daily variables that corresponded to DAYMET, in the following CMIP6 models: CanESM5, ACCESS-ESM1-5, MIROC6, MPI-ESM1-2-LR. Separate LSTMs were tuned and trained using the historical simulation data from these CMIP6 models. These models were able to predict the annual breakup event within an average of 7.7 days. The optimal model architecture derived from the historical simulation was used to predict the breakup event under a variety of experiment scenarios for each of the CMIP6 models (ie ssp119, ssp245, ssp370, ssp585 and ssp534-over). Results of these scenarios show that as concentrations of atmospheric CO2 increase, leading to hotter global temperatures in the Arctic region, the annual breakup events tend to happen earlier and earlier in the spring season. Results from our study provide insights on changes that may happen to breakup events in rivers across Arctic Alaska under various climate change scenarios.

Category
Cryosphere
Funding Program Area(s)