Deep learning techniques give scientists the longest–lead time forecasts yet.
By Jenessa Duncombe
The artificial intelligence technique deep learning is everywhere in our daily lives if you know where to look. Siri’s voice commands, online banking, and photo tagging on social media all use deep learning to uncover powerful structures hidden in data.
Yoo-Geun Ham can now add another item to the list: El Niño forecasting.
Ham and his collaborators created a model using deep learning that forecasts El Niño and La Niña events 18 months in advance, beating current models that forecast only 1 year ahead. Using simulated data from a climate model, they trained their data-driven model to sidestep barriers faced by many contemporary models, resulting in what Ham called “the world’s best” El Niño forecasting model.
Forecasting El Niño and La Niña events could help regions better manage food prices, disease, and water shortages, but predicting when an event will occur is challenging. Researchers have poured decades of study into understanding the global climate phenomenon that drives El Niño and La Niña events, the El Niño–Southern Oscillation (ENSO). Despite major advancements, scientists cannot predict events more than 1 year away, even though the physical precursors to El Niño and La Niña, such as shifting ocean temperatures, may occur more than a year in advance.
Many state-of-the-art forecasting models use mathematical equations to power their predictions. These models elegantly simulate the physical relationships between the ocean and the atmosphere, but they contain slight errors that compound over time, rendering long-term forecasting unmanageable. On the other hand, models based solely on analyzing data, called statistical models, have traditionally lacked a sufficient number of measurements to make them robust.
The latest study led by Ham from Chonnam National University in South Korea built a statistical model that skirts the problem of data scarcity. Ham and his colleagues fed their model data from both sophisticated climate models and an ocean reanalysis model that gave them global snapshots of ocean temperatures since the late 19th century. Using these model outputs increased the number of available data from about 150 measurements to nearly 3,000 per month.
With a new trove of available data, the scientists used an artificial intelligence technique called deep learning to analyze it. Deep learning is often used in image recognition: The technique identifies noteworthy qualities in an image and systematically classifies it through a series of steps. For example, a deep learning model will discern a cluster of pixels denoting an “edge” in an image, such as the black and gray edge of a cat’s ear, and through a series of steps determine whether that edge belongs to a recognizable object, like the head of a British shorthair cat.
Ham and his colleagues used a similar technique: For each global snapshot of ocean temperatures from the climate model, they taught their deep learning model to identify telltale signs that an ENSO event was on its way, such as abnormally warm water in the Indian Ocean or tropical Pacific. By feeding the model snapshots from 1871 to 1973, as well as telling it the correct answer for when an ENSO event occurred, they trained the model to make future forecasts.
To test their model, they fed it global snapshots from 1984 to 2017 but withheld the answers of when ENSO events occurred. Their model successfully forecasted events 1.5 years in advance and nicely predicted an event’s amplitude. The new method outperformed eight other current forecasting models and was even able to predict the specific subclasses of El Niño and La Niña. The researchers published their results in Nature on 18 September 2019.
Ham said that this technique works even though the climate model seeding their statistical model contains errors. The deep learning technique corrected the systematic errors, he said, for reasons the researchers do not fully understand.
“This looks like breakthrough work,” Youmin Tang, a professor of environmental science and engineering at the University of Northern British Columbia, told Eos. The study “may excite a new round of application of machine learning on climate predictions.”
Michael Tippett, an associate professor of applied mathematics at Columbia University, agreed that it was an exciting development. “I hope we will see these forecasts used operationally to make predictions about the future so that their practical skill can be assessed.” Ham now posts future ENSO predictions using this method on his research website.
Ham said he and his colleagues are working with artificial intelligence experts to optimize their deep learning technique. When asked how far future models using deep learning may forecast, he said, “I’m not sure what is the upper limit.”
Ham said he plans to apply the same methodology to other climate signals, which would be “very easy to do in its current stage.” He said forecasting the Indian Ocean Dipole, which affects the monsoons in India, might be next.