Science1 publisherNot yet confirmed elsewhere2 min readPublished
Explainable AI links Great Lakes water levels to weather from the season before
Researchers trained eight algorithms on 40 years of Great Lakes data and found the inputs' influence on lake levels grows after three or four months. That lag points to months of lead time for dam operators and flood mappers, if it holds in forecasts made before the fact.
The Scientist · Science desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- Lake Michigan went from a record low in January 2013 to a record high in summer 2020, nearly two meters apart, and Lakes Superior, Erie and Ontario reversed within a few months of it.
- The standard water-balance method adds inflows and subtracts losses, but the authors say it needs extensive calibration and misses unusual climate variations.
- To explain the models, the team split each month's rise or fall among the inputs with SHAP values, then perturbed the models with a method called VARS to measure time lags.
- The drivers differed by lake: runoff and outflow for snowmelt-fed Superior and Michigan, upstream inflow via the Detroit River for shallow Erie, and evaporation for Ontario.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- capability Because the model ties June levels to January snowfall, a forecast built this way would have months of lead time for harbor, intake and shoreline planning, if the lag holds outside the training years.
- constraint With both records inside the training window, the reported accuracy cannot yet support a claim that the 2013-2020 reversal was predictable from earlier data.
- decision Each lake needs its own attribution before it guides a dam release or a flood map, since one lake's driver profile does not transfer to the next.
Each of the nine inputs went into the models seven times: once for the current month and once for each of the six months before it, 63 inputs per algorithm in all [9][19]. That design is what let a model tie a June water level to January snowfall [9]. The authors expected the older inputs to fade, the way a downpour stops showing up in a river after a few days [14]. At lags of three or four months they found the reverse [14]. "A lake does not react to yesterday's weather, but to the previous season's," the authors wrote [15].
The denominator is modest. Forty years of monthly data comes to about 480 values per lake, set against those 63 inputs [20][19]. The four best of the eight algorithms came within about 12 centimeters of measured levels, against 14 to 21 centimeters for the simplest models [1]. The gain is 2 to 9 centimeters, roughly 14% to 43% depending on which simple model is the baseline [17]. Set against a swing of nearly two meters, a 12-centimeter error is about 6% of the move [18].
The authors write that almost no one saw the 2013-to-2020 reversal coming [4]. The thing this doesn't tell you is whether their models would have. The 1982-to-2022 training window contains both the January 2013 low and the summer 2020 high [21]. The authors' account does not say which months were held back to measure the errors [1]. Even a random test split inside that window would leave the models trained on months from the reversal itself [21].
Both explanation tools question the trained model [10][11]. SHAP divides the model's own output among its inputs, and VARS perturbs the model, not the watershed [11]. Their answers describe what a model learned from 40 years of data, and a model can lean on a variable that only moves alongside the real cause. The physical check is whether the attributions match how each basin works. With the same method and the same variables, the four lakes produced four different results [13], and the authors tie each one to the lake's physical makeup [12].
The authors' case for explanation is a practical one. A dam operator or a flood-zone mapper needs to know why a level is moving, along with the number, they argue [7]. Those levels set harbor depths, shoreline stability and drinking-water intakes, and decide whether coastal wetlands survive [16]. In my view the lag finding is the result to test first, on lake levels from after 2022, the last year in the training data [8].
What to watch
- A hindcast that trains the same algorithms only on years before 2013 and tests whether they foresee the climb to the 2020 record.
- Whether the three-to-four-month lag peak reappears when SHAP and VARS are run on a physically based water-balance model of the same lakes.