AI could predict droughts before they strike, Virginia Tech expert says - WSLS

BLACKSBURG, Va. – More than 100 million people across the Midwest and Southeastern United States are dealing with drought conditions that are threatening crops and livestock. The typical response is reactive — voluntary water conservation measures after dry conditions have already taken hold. But a Virginia Tech researcher says artificial intelligence could change that, if the models have enough data to work with.

Craig Ramseyer, an associate professor in Virginia Tech’s Department of Geography, says AI is already showing promise in detecting early drought signals — but the technology is only as good as the data feeding it.

“We’re collecting more weather and climate data than we ever have before,” Ramseyer said. “And so the AI can start to learn complex patterns that maybe we haven’t been able to detect in the past.”

Droughts are arriving faster than before

One of the biggest challenges researchers face is what Ramseyer calls “flash drought” — conditions that can shift from wet to dry within weeks rather than months.

“In the last couple of decades, one of our concerns with the drought is that we enter drought more rapidly than we used to,” he said. “We’ve seen this recently with Smith Mountain Lake levels really dropping, because the atmosphere is really thirsty, and we’re not replenishing much rain, so it just draws that into the atmosphere through evaporation.”

That speed makes early detection critical. Wildfires, water shortages, and agricultural failures are often the result of droughts — and by the time conditions are identified, it can be too late to prepare.

What AI can — and can’t — do

Ramseyer says AI models excel at one thing in particular: recognizing patterns from past droughts and applying them to current atmospheric conditions.

“Once we’ve experienced the drought, we can go back in time and figure out what the fingerprint was,” he said. “The more times we feed the AI models with examples of what produced a drought, the better it’ll get at saying, hey, three weeks from now, a similar pattern looks like it’s setting up across the global atmosphere.”

But AI has a significant limitation — it cannot predict what it has never seen before. Ramseyer pointed to Hurricane Helene as an example of where traditional physics-based models have the edge.

“The ability to mathematically tell us that Helene was going to dump more rainfall than we’ve ever seen before — that’s not a strength of AI models, because it’s got to have already seen the data before,” he said. “It’s got to have already been able to detect the patterns.”

That’s why major weather modeling agencies are now running AI models alongside traditional mathematical, physics-driven models — using each for what it does best.

The data gap problem

The most significant barrier to AI-powered drought forecasting isn’t the technology — it’s the shortage of quality data. Ramseyer compared weather data to what powers tools like ChatGPT.

“We don’t have near that amount of data for weather or for soil moisture and things happening on the land surface,” he said. “So we have a long way to go because the AI models are only as good as the data that we’re able to put into it.”

Data gaps exist both domestically and internationally — and they extend far above the ground. Weather systems can reach 50,000 feet into the atmosphere, but current measurement tools don’t come close to capturing that full picture.

“How many measurements are we making right now at 40,000 feet above the surface? Not many,” Ramseyer said. “In fact, that’s what we would launch these weather balloons for, but we only can launch them twice a day right now. So we’re getting two snapshots at like 50 locations across the U.S. of the whole depth of the atmosphere.”

Weather patterns affecting the U.S. can also originate from as far away as Asia, and incomplete sensor networks anywhere in the world introduce uncertainty into every model — AI or traditional.

“We need great U.S.-based sensors, but it’s really a global issue,” Ramseyer said. “Other countries have much bigger gaps. And all of that leads to uncertainty in our models.”

He added that forecast failures — like a snowfall prediction that doesn’t pan out — often come down to data quality, not the models themselves.

“If you’re mad at your meteorologist because the snowfall forecast didn’t pan out, the real issue actually is the data we had to put into the model,” Ramseyer said. “AI or not, we need more data to improve your forecasts.”

What a ‘drought buster’ actually means — and why they’re harder to find

Even if AI helps predict a drought, ending one requires the right kind of rain. Ramseyer says climate scientists informally refer to the ideal recovery event as a “drought buster” — and it’s becoming increasingly rare.

“A drought buster basically means a large widespread rainfall event that’s long duration, not too extreme on precipitation, not too little,” he said. “If we get too much rain too quickly, we can’t actually capture a lot of that water down into the lower parts of the soils, and a lot of it ends up running off and causing erosion and getting rapidly into the rivers and streams and causing flash flooding.”

The ideal event, he says, is steady and moderate — the kind of slow, all-day rain that soaks into the ground rather than rushing off into waterways.

“It’s kind of that nice long duration, hard to get out of bed day where it just kind of rains all day steadily,” Ramseyer said. “That would be the ideal kind of event, but those are harder to get now than they used to be.”

What comes next

Ramseyer says the path forward combines more satellites, more rain gauges, faster data collection intervals, and an expanded global sensor network — all feeding both AI and traditional models working in tandem.

“We really need data, traditional models and AI models all kind of working together to improve this,” he said.