The frozen lake is now being asked to keep its own thaw calendar
the frozen lake is now being asked to keep its own thaw calendar: spatial temporal transformers forecast ice cover extent across five large canadian lakes, a random forest reads sentinel 1 radar to time ice on and ice off across 30,862 alaska thermokarst lakes at 94 percent accuracy, and a deep net predicts river ice breakup on eight alaskan rivers.
the lake now files its own freeze ledger.
Context
LIF-DL (EGUsphere preprint) uses Spatial-Temporal Transformer Networks to forecast lake ice cover extent across whole lake surfaces for Great Slave Lake, Great Bear Lake, Lake Winnipeg, Lake Athabasca and Reindeer Lake. It trains on 2004 to 2017, makes one-week forecasts, and is run autoregressively over a 2018 to 2022 holdout. Freeze-up and break-up are predicted within 3 to 9 days of observation against 5 to 22 days for the FLake physical model.
The Alaska North Slope study (EGUsphere preprint) classifies Sentinel-1 radar with a Random Forest trained on Sentinel-2 imagery and ERA5 temperature, for 2019 to 2023. It reports 94% accuracy for ice state across 30,862 lakes, a mean ice-free period of 115 days, and ice-free duration shortening by about 30 days per degree of latitude northward.
The river ice study (Water Resources Research, 2025) uses a Long Short-Term Memory model on 33 locations along eight rivers in Alaska and Western Canada. It reports a mean absolute error of 5.40 days for breakup date, and with seasonal forecast ensembles the predicted breakup falls within 5 to 14 days of observed.
The three studies predict different things. LIF-DL forecasts ice cover over a week on five large lakes; its 3 to 9 day figure is timing error for freeze-up and break-up events. The Alaska work is a classification of ice state from radar over 30,862 lakes, a measurement and not a forecast, and its 94% is accuracy for ice versus open water. The river model forecasts one date per site per year, using weather data and watershed attributes. The note says 'a deep net' for rivers; the paper uses an LSTM, and the 33 locations include Western Canada. Both lake papers are preprints under review. Simple degree-day methods are also strong baselines for timing, and the Alaska paper reports its Random Forest timing agrees with an accumulated degree-day model within two Sentinel-1 repeats for about 60% of lakes.
Related work
- A Deep Learning Approach for Lake Ice Cover Forecasting (EGUsphere preprint) ↗The LIF-DL paper with the five Canadian lakes and the FLake comparison.
- Multi-sensor satellite analysis reveals latitudinal and morphometric controls on ice phenology across 31,000 thermokarst lakes on the Alaska North Slope (EGUsphere preprint) ↗The source for 30,862 lakes and the 94% accuracy.
- Long Short-Term Memory Model to Forecast River Ice Breakup Throughout Alaska USA (Water Resources Research) ↗The river ice breakup LSTM with the 33 locations and 5.40 day error.
Watch next
- Look for peer-reviewed versions of the two lake preprints and for LIF-DL results on lakes outside the five training lakes. For rivers, check the LSTM against operational breakup outlooks.
Sources
- A Deep Learning Approach for Lake Ice Cover Forecasting (EGUsphere preprint)egusphere.copernicus.org
- Multi-sensor satellite analysis reveals latitudinal and morphometric controls on ice phenology across 31,000 thermokarst lakes on the Alaska North Slope (EGUsphere preprint)egusphere.copernicus.org
- Long Short-Term Memory Model to Forecast River Ice Breakup Throughout Alaska USA (Water Resources Research)agupubs.onlinelibrary.wiley.com
Provenance
The note above is reproduced unedited from the original post, first published on Threads on 4 October 2026 at 03:01 IST. Sources are the papers and datasets the note draws on.
View the original post ↗Embed this note
More notes
The air is now being asked to keep its own ledger
the air is now being asked to keep its own ledger: ecmwf’s aifs compo becomes the first ai model to forecast atmospheric composition globally every three hours, cleanair simulates 365 days of pm2.5 over china in ten seconds, and a unified framework maps six pollutants at one kilometer across the whole country. the air now files its own composition report.
read the note →The current is now being asked to draw its own map
the current is now being asked to draw its own map: china’s langya 2.0 predicts six ocean phenomena including internal waves and mesoscale eddies, a deep net called wenhai resolves eddies globally with air sea flux formulas built in, and scripps infers surface currents from the way temperature patterns deform in satellite images. the ocean now files its own circulation report.
read the note →The soil is now being asked to report its own carbon
the soil is now being asked to report its own carbon: a nix color sensor paired with generative data augmentation predicts soil organic carbon without a lab, random forest drives 74 percent of soil health mapping studies, and sentinel 2 tracks five year carbon change across france and italy from 922 samples. the dirt now files its own carbon account.
read the note →