Thermal-time models fitted to 1960–1990 air temperature predict 44–64% of the later lake ice freeze-up in 1991–2026, and 73–84% of the earlier break-up, at lakes in Finland and the LIAG database

We tested whether lake ice models that use only air temperature continue to work in a warmer climate. For each lake, we fitted a thermal-time (degree-day) model to ice dates and daily station air temperature of 1960–1990. Then we predicted the 1991+ dates from the measured temperature. We used 55 Finnish lakes (SYKE records to 2026, FMI stations) and up to 69 lakes from the NSIDC LIAG database outside Finland (mostly USA, records to 2011–2019, GHCN-Daily stations). Freeze-up became 10.4 days later in Finland and 7.0 days later at the LIAG lakes. The models predicted only 50% and 56% of this change; the mean bias (observed minus predicted) was +3.9 days (95% year-bootstrap interval +1.8 to +6.2) and +2.8 days (+1.6 to +4.0). Finnish break-up became 6.9 days earlier; the model predicted 76% (bias −1.5 days, −2.3 to −0.7; earlier than predicted at 47 of 55 lakes). In Finland, the residuals became larger with time. Models fitted to 1991+ gave the opposite bias for 1960–1990, and random-half splits gave a bias near zero. Synthetic dates with a fixed temperature–ice relation were recovered at 91–102% with a bias of less than 0.6 days, also with a second station. Thus the fitting method and station noise do not explain the residual. A linear model, models with winter or summer temperature terms, and a net degree-day model also under-predicted the change (41–81%). At the LIAG lakes, raw GHCN-D stations warmed about half as much as CRU TS in spring. With CRU-adjusted temperature, the model predicted 84% of the LIAG break-up advance (bias −0.4, −1.3 to +0.5), so most of that residual was a station artifact. The Finnish results did not change (break-up 73%, freeze-up 44%), and a LIAG freeze-up residual remained (64%; +1.9 days, +0.6 to +3.3). Our break-up results agree with a parallel SCI study that used gridded temperature. Air-temperature models probably under-estimate the future delay of freeze-up. Limitations: one station per lake; no data on lake depth, snow, radiation or water temperature, so we cannot identify the cause; possible changes in observation practice; large freeze-up model errors (RMSE 10–14 days).

lake ice · ice phenology · freeze-up · break-up · ice-off · degree-day model · thermal time · out-of-sample test · non-stationarity · Finland · SYKE · FMI · NSIDC G01377 · GHCN-Daily · CRU TS · station inhomogeneity · climate change