Adding large-scale climate patterns to algae-prediction models could lead to earlier and more accurate seasonal forecasts, says a new IBM Research study.
Toxic algae blooms are on the rise worldwide, fueled by increased fertilizer use and a growing human population. Harmful to people and animals alike, they are now endemic on Lake Erie, the smallest and shallowest of North America’s five Great Lakes. Fallout from the blooms has an annual cost of at least $272 million to shoreline communities in Ohio, Michigan, and Canada’s Ontario province.
The solution to Lake Erie’s water woes involves cutting the flow of nutrients to the lake from agricultural lands and aging sewage treatment plants. But in the short-term, advanced warning of a bad blue-green algae outbreak can give communities valuable time to prepare. Boating and swimming can be moved to less polluted areas. Public drinking water can be filtered and treated before anyone gets sick.
The U.S. National Oceanic and Atmospheric Administration (NOAA) puts out a forecast each year in late June, and though generally reliable, its models underestimated the relatively severe blooms of the last two summers. NOAA’s forecasts also missed the record-setting outbreaks of 2011 and 2015 when a thick layer of green scum covered the lake for about 2,000 square miles.
NOAA’s forecasting program began in 2012 with a model based on phosphorous loading to the lake and has since expanded to include temperature data and estimates of spring rain and snowfall. But its failure to capture severe outbreaks led IBM researchers to wonder if other variables might be contributing. They turned to AI to see if a A genetic algorithm finds the best possible solution to complicated problems in a similar way to Darwin’s theory of evolution. In a series of simulations, the algorithm identifies the fittest solutions and allows them to reproduce and iteratively improve until the optimal solution emerges. Genetic algorithms have been used to a range of problems from solve sudoku puzzles to optimizing decision trees.genetic algorithm could pick out patterns in the data that humans had missed.
In a study in Nature Communications Earth and Environment, the researchers identify El Niño and a pair of related climate patterns as top predictors, a finding that could potentially lead to improved forecasting. “We know that large-scale climate patterns impact weather over Lake Erie and the hydrodynamics of the lake itself,” said the study co-author, Lloyd Treinish, a climate scientist at IBM Research. “So, it makes sense they may also drive algae growth: More precipitation to this region in spring means more nutrient-rich runoff to the lake, feeding potential toxic algal blooms.”
If IBM’s work is validated, and NOAA incorporates largescale climate data into its forecasts, policymakers could potentially gain an extra month of lead time, as climate signals show up earlier, the researchers said. Accuracy could also improve. IBM’s model based solely on climate data retroactively predicted that last summer’s outbreak would be a moderately-severe 6 to 8 out of 10 on NOAA’s bloom-severity scale. NOAA, by contrast, predicted in late June a relatively mild bloom of 3.5. The actual bloom, at the end of August, came in at 6.8.
Responding to IBM’s results, NOAA researchers said the study was interesting, but argued that large-scale climate patterns are already effectively captured in its models through temperature data, as well as rain and snowfall estimates, in its spring forecasts. El Niño’s influence on weather in the northern U.S. is also less pronounced than it is in the south, they said, potentially limiting its predictiveness for Lake Erie.
“There is a lot of uncertainty in modeling precipitation out 1-2 months,” said Richard Stumpf, an oceanographer at NOAA who leads the Lake Erie forecasts. “We will examine any model that may help us do so more accurately.”
The link between cyclic warming in the Pacific Ocean and blue algae blooms on Lake Erie
El Niño, known formally as the El Niño Southern Oscillation (ENSO), is a climate pattern that starts in the tropical Pacific but wreaks havoc on weather worldwide. Every few years, and sometimes longer, the eastern Pacific gets warmer than usual, shifting the westerly winds, which changes rain and snowfall patterns across the planet.
In the northern U.S. and Canada, El Niño years are typically warmer and drier than normal while La Niña years are the opposite; temperatures get cooler, and more rain and snow fall up north, including on Lake Erie. ENSO’s effect on Lake Erie is well documented, which is why IBM researchers thought it might be a strong predictor for toxic blooms.
IBM also wanted to see if nutrients other than phosphorous might play a role. The Clean Water Act in 1972 drastically improved water quality in Lake Erie by forcing communities to build sewage treatment plants to limit the flow of raw sewage to the lake. But two decades later, those gains were reversed by increased fertilizer and manure runoff, providing a nourishing stream of nutrients for blue-green algae.
Both U.S. and Canada have agreed to reduce their phosphorous output, but the pact is voluntary. There is also concern that other nutrients might be feeding the blooms, including nitrogen from sewage treatment plants and chlorides from winter road salt.
To evaluate these other factors, IBM researchers set a genetic algorithm loose on Lake Erie’s historic toxic bloom data and a list of possible contributors to see if the AI might find connections too subtle for humans to detect. In the end, the algorithm picked phosphorous over seven other nutrients, and three major patterns from a list of climate patterns — the November phase of the Pacific Decadal Oscillation (PDO), December phase of the Pacific North American (PNA), and April phase of ENSO.
When they ran their model based on the phosphorous and climate data, they found it could explain 60% of the variation between the severity of the predicted blooms and what actually happened, versus 45% for the phosphorous-only model, based loosely on NOAA’s original prediction model from 2012.
The researchers next analyzed the last two decades of local weather data to see if wetter seasons lined up with worse outbreaks. The connection was striking.
Years with mild algal blooms, they found, were linked to ENSO’s El Niño phase and positive phases of the PDO and PNA in winter, while years with severe blooms were connected to La Niña, which tends to bring cold air down from the Arctic, lowering temperatures and triggering more snow and rainfall in the northern U.S.
“The observational data provided additional validation, and also helped us build several models to predict the severity of toxic bloom on Lake Erie,” said the study’s first author, Mukul Tewari, an expert on climate prediction at IBM Research.
It’s too soon to tell what the summer of 2023 holds for Lake Erie, but stay tuned for May, the researchers said. By then, enough climate data will have come in to say how bad communities can expect the blooms to be.
- Note 1: A genetic algorithm finds the best possible solution to complicated problems in a similar way to Darwin’s theory of evolution. In a series of simulations, the algorithm identifies the fittest solutions and allows them to reproduce and iteratively improve until the optimal solution emerges. Genetic algorithms have been used to a range of problems from solve sudoku puzzles to optimizing decision trees. ↩︎