All-weather lake-ice phenology from Sentinel-1 and reanalysis thermodynamics: a validated, transferable method applied to 131 Adirondack lakes

Lake ice phenology is a sensitive and physically interpretable indicator of climate change, but observer records exist for only a small and shrinking subset of lakes, and optical satellite methods are limited by cloud in precisely the season when freeze-up and break-up occur. We present a method that estimates freeze-up and break-up dates by combining Sentinel-1 C-band backscatter with gridded reanalysis thermodynamics in a single probabilistic model. Per-scene radar features, cumulative freezing and thawing degree-days, seasonal harmonics and static lake morphometry enter a ridge logistic regression model that returns the probability of ice for each scene. A two-state hidden Markov model with degree-day-gated transitions then converts each season's probability sequence into freeze-up and break-up dates, each with a resolution bound of half the revisit interval that brackets it. Under leave-one-lake-out cross-validation across 13 lakes and 79 lake-seasons, the method resolves freeze-up to a mean absolute error of 11.5 d and break-up to 8.3 d, against 36.1 d and 23.4 d for a fixed-threshold baseline that additionally fails to resolve break-up in 89 % of seasons. Withholding each information source in turn shows that the radar contribution is concentrated at freeze-up, where it reduces error by 4.6 d, and is negligible at break-up, which cumulative thawing degree-days already constrain. Combining the two sources also suppresses false ice detection: a meteorology-only model declares ice in all four documented no-freeze winters on Lake George and the fused model in one. Applying the developed model to 131 Adirondack lakes over ten winters yields 1,310 lake-season records with a resolution bound attached to every event, recovering an elevation gradient of 6.0 d per 100 m and a 36 d difference in median ice duration between the coldest and warmest winters. Because the method requires no local observations and no per-lake calibration, it provides a consistent baseline for regions where ice records are absent, and offers a route to extending ice-phenology monitoring to the large majority of lakes that have never been observed.