Lake ice-on dates in 1991-2020 were 3.8 days later than 1901-1990 autumn air-temperature models predicted at 73 northern lakes; ice-off dates were 1.2 days earlier than predicted at 151 sites

We tested if relations between air temperature and lake ice dates from 1901-1990 stay correct in 1991-2020. We used NSIDC G01377 lake ice records and Berkeley Earth monthly air temperature. For each site we fitted the models on 1901-1990 only and predicted 1991-2020 dates from the observed temperature. At 73 lakes (mostly Finland and the United States), ice-on was 7.3 days later than the 1901-1990 mean. Autumn temperature predicted half of this delay. Ice-on was 3.8 days later than predicted (95% year-bootstrap interval 2.2 to 5.4). 70 of 73 lakes had a positive residual, and the residual increased by 3.2 days per decade in the test years. Threshold-crossing and quadratic temperature models, summer temperature, the same-year ice-off date, lake depth and GHCN-CAMS temperature did not remove it. Placebo splits before 1991 gave smaller residuals (+0.6 and +1.5 days, intervals include zero). At 151 sites, ice-off was 4.6 days earlier than the mean. Spring temperature predicted most of this. The residual was -1.2 days (-2.2 to -0.2). With winter temperature added, it was -0.7 days (-1.5 to +0.1). Sensitivity in days per degree was larger in warm years, as the slope of the seasonal temperature curve predicts, but this explains only about 0.5 days. Limitations: few North American and almost no river sites, monthly 1-degree temperature, excluded ice-free seasons, and follow-up tests specified after the primary result. Code, design, aggregate results and figures are included. Source data stay outside the package.

lake ice · ice phenology · freeze-up · break-up · air temperature sensitivity · out-of-sample test · NSIDC G01377 · Berkeley Earth · climate change · nonstationarity · limnology