LAKE GEORGE, N.Y. (NEWS10) – The weather hasn’t quite been itself in the last week of December or first days of January, with temperatures routinely above freezing leading to less snow on the ground than one might expect from a typical North Country winter. In the village of Lake George, there’s one other, 32-mile-long way of telling how wintery the season has really been. All you need to do is find out whether the lake is frozen yet.
Determining when the lake freezes is a lot more complicated than just tracking how cold the air is; although that’s not a bad place to start. Researchers at The Jefferson Project – a collaboration between IBM, Rensselaer Polytechnic Institute and the Lake George Association – operate an array of over 50 sensors and survey stations across the lake, and the surrounding tributaries that feed it. The data collected there can be used for a lot of things, including to estimate when the blue surface of the water will turn to hard ice.
In 2020, some researchers at the project did exactly that, in the published article “Predicting complete winter ice coverage at Lake George, New York.” The article, written by IBM researchers Campbell D Watson, Guillaume Auger, Mukul Tewari and Lloyd Treinish, as well as former RPI researcher Kenneth Johnston, explains some of the factors that go into determining when the lake will freeze, and also provides some insight into how things may change in the future.
“Ice records have been maintained at Lake George since 1912 and, other than in 1919, it completely froze over each winter until 1990,” say researchers in the article’s introduction. “Since 1990, Lake George has failed to completely freeze over 13 times.”
There are a few different methods to getting a grasp on how and when a water body will freeze. Some require data on parameters and physical makeup of a water body that aren’t currently available for Lake George at the level researchers would need. Instead, the article uses data from local weather, a history of Lake George ice records that date back to 1912, as well as machine learning processes that analyze available points of data in order to find what information about the lake and its surrounding environment are most likely to be helpful.
It’s important to note that weather patterns are consistently on the move. The machine learning programs at work in the study were fed data on a set of different weather variables that factor into freezing. Those variables include air temperature; wind speed; the presence or absence of snow and precipitation; cloud cover; time of year, measured in days past Sept. 1 of a given year; relative humidity, in percentage; and surface pressure in the water itself. The data came from a range of years, from 1974 to 2015, and was sourced from Floyd Bennett Memorial Airport in Queensbury.
The researchers also ran experiments that only used the data on air temperature to predict freezing. When they did so, the accuracy of their predictions fell from 90-91% to 80-85%.
“This supports the notion that ice-in and ice-out is sensitive to more weather variables than just air temperature,” the article reads.
Some of those variables can be tracked by anyone interested in real-time at the Jefferson Project’s live data map. For example, on Monday afternoon, the project’s Cedar Lane weather station near Million Dollar Beach in the village of Lake George was tracking an air temperature of 16 degrees Fahrenheit; a north wind clocking in at 9 miles per hour; and 86.2% humidity. On Tea Island, in the middle of the lake and just south of Plum Point, a vertical profiler at the same height tracked an air temperature of 30 degrees; a water temperature of 52 degrees; and a 5 mph southbound wind.
Not every station is always operating. On Monday, none of the sensors in the lake’s upper miles or in Ticonderoga were feeding live data. There’s a range of what different locations track, from the master list of important factors.
In the 2020 paper, each one of those factors is then split into its own set of averages. The program looks at the data on each variable over spans of 1, 3, 7 and 14 days, as well as the previous 40 weeks, divided up into sets of four.
The researchers fed those factors into the program in sets, dating from 1974 all the way to 2015, and generated two graphs that show how close a given year gets to 100% likelihood of full ice cover across Lake George. The first one, spanning from 1974 to 2000, was considered a “training period” for the software to learn off of.
As for the 2001-2015 data, the program accurately predicted the timeframes for ice during 11 out of 14 winters. Eight correct predictions were among the 10 winters when Lake George did entirely freeze, and the other three were among four winters when it didn’t.
Imperfect results don’t mean faulty programming; in fact, the researchers feel good about what the predictions could mean when looking into the lake’s future. Machine learning means that the software that made the 1974-2015 predictions will be all the more well-equipped for what’s next.
“An attractive prospect following the results above is to use the same classification techniques with climate model projection data,” the report says. “The same trained models are used to make projections of ice coverage through the 21st century using downscaled climate model data.”
The future projected when using the same kinds of data paints a grim picture, where the impact of climate change may very well make regular, reliable ice on Lake George a thing of the past. Scenarios generated in the report only show some years as having a completely frozen lake between 2020 and 2040. One set of projections extends from 2040 to 2060, and doesn’t show a single year of complete ice cover on the lake.
More recently, the Lake George Association reports that Lake George froze only briefly in 2016-17. When facing these potential futures, the researchers behind the study have some words of clarification, as well as a reminder that any amount of data can always be made more precise.
“While these future projections of complete ice coverage at Lake George are grim, we remind the reader that these are projections for whether one to traverse the entire length of the lake without crossing open water, not for the presence of lake ice,” the report reads. “Also note that have used data from only one global climate model and CMIP5 had over 40 global climate models participate in the experiment. Extending the analysis to include an ensemble of downscaled global climate models would provide much greater certainty on projected temperature changes and the loss of complete ice coverage at Lake George.”
The Jefferson Project’s active surveying of Lake George can be seen online in live feed form. In addition to freezing, it also tracks data on salt, water depth, temperature and the quality of water entering the lake from tributaries.