From sudden floods to weeks of scorching heat, increasingly unstable weather is a headache for U.S. farmers. Known as “weather whiplash,” these disorienting swings between too much rain and not enough — from inundations to droughts — are accelerating thanks to climate change. Soil moisture levels, often overlooked, are highly susceptible to these vacillations. But farmers, who are already suffering the consequences of increasingly unpredictable weather patterns, are all too aware of the issue — because it’s right there under their boots.

As the world warms and weather patterns become less stable, soil moisture increasingly fluctuates. Farmers are finding soil moisture levels more difficult to predict. And the stakes are high, because getting it wrong can be devastating. One mistake can lead to overwatering, or not irrigating enough, or planting at the wrong time, risking significant crop losses.

But now, thanks to artificial intelligence (AI) tools developed by CALS researchers, help is on the horizon.

–-–

In 2011, U.S. cereal yields declined, and droughts and floods were behind 70% of the downturn. Eight years later, floods across the Midwest dramatically reduced agricultural production. Researchers estimated that an unusually sodden spring had increased soil moisture, leading to a 10% crop loss in parts of America’s heartland. One explanation: an uptick in soil moisture levels that had taken farmers by surprise.

When it comes to soil moisture, farmers are often in the dark. “Some farmers rely on the weather and the weather forecast to guess what the soil moisture level will be,” says Jingyi Huang, an associate professor of soil science. “For example, if it rained yesterday, a farmer might guess the moisture is very high here in the field. The problem is, they don’t always know how fast their soil will dry out, particularly when crops are growing in the field. And if they have a large farm, one side of the farm can be totally different from the other.”

But what about testing the soil the old-fashioned way — driving out to a field and checking how damp the soil is by crumbling it in your fingers? All well and good, says Huang. But what if you have 50 fields?

“Are you going to drive out to each one and dig holes to feel the moisture level not only at the surface but also at depth where plant roots can access? And in any case,” he says, “soil moisture can vary across a single field and with depth, and some fields are in the hundreds of acres.”

There are probes farmers can place in the ground to detect soil moisture at depth, but they are very expensive. And their usefulness is limited because a large field may contain a lot of variation, requiring many sensors — sometimes hundreds of them.

“Next time you take an airplane, look down,” says Huang. “Fields next to each other will look different, some lighter and some darker, and sometimes a single field will contain different shades of brown. This is most likely due to different wetting and drying patterns.”

Accurate knowledge about the water content of the soil is important for other reasons too. More precise soil moisture maps will allow researchers to more accurately predict droughts, floods, and wildfires. And soil moisture tells us about how the climate is changing. Better soil maps mean more accurate global climate models.

“If you do not properly model soil moisture, then a lot of climate-change forecasts will not be accurate,” says Huang. “With a better understanding of soil moisture, we will have a clearer understanding of climate change and its impacts.”

Soil moisture is inherently complex, and it frustrates scientists even at the best of times. The soil environment is an intricate ecosystem, with many variables driving moisture levels.

“Soil moisture estimation is complex due to various influencing factors, like soil types, precipitation, and vegetation,” says Yijia Xu, a biological systems engineering Ph.D. student studying digital agriculture. (Xu works in the lab of one of Huang’s collaborators, associate professor Zhou Zhang, but assisted Huang with the project.) Throw in the disruptive effects of climate change, and researchers have long feared that accurately mapping and forecasting soil moisture is out of reach.

But Huang’s team at CALS is starting to unearth solutions. Harnessing the power of AI and satellite data, his lab is building sophisticated new tools that will empower farmers to prepare for a less certain world.

Huang’s work, which is supported by the U.S. Department of Agriculture, will save farmers time and money, keep the nation fed, and shed light on how climate change is reshaping our world. “Our project is going to deliver a soil moisture map with an extremely high spatial and temporal resolution,” says Huang. “We are using AI and satellite data to gather this information and give it to farmers.”

–-–

The roots of Huang’s new and game changing tech go deep. In 2013, when Huang arrived in Australia to begin his graduate studies, he was struck by the variation in the soil’s color and texture. “It’s amazing to see the differences in soil when you travel,” he says.

At the University of South Wales in Sydney, Huang began mapping these variations with handheld geophysical instruments, including an electromagnetic induction device that works like an MRI scanner. He used it to read below the surface to the root zone, around 60 inches down. His mission: understanding how salts were moving through the soil.

Huang moved to CALS after earning his Ph.D., and it was in Wisconsin that he started to focus on soil moisture. Realizing that there was little high-resolution soil moisture data, he began thinking about how to measure soil moisture at an unprecedented level of detail. The ground-based proximal sensing devices he had been using to monitor salt were not up to the task. To provide useful soil moisture maps, he would need to somehow scale up the sensing process dramatically.

The key, he soon realized, to understanding soil lay not in the ground but in space. He began gathering satellite images of fields across the U.S. But he quickly ran into problems with remote sensing too.

“All the current satellite data, including what NASA produced, had a low spatial resolution for soil moisture. The pixel size was around several miles, which is not fine-grained enough to be useful,” says Huang. “And although the moisture of the top two inches of the soil’s surface can be sensed by satellite data, farmers need data about what is going on at deeper depths, in the root zone, up to 60 inches underground.”

The traditional way of modeling soil moisture uses physics-based process models. These complex models are mathematical representations of the physical processes that take place in the soil environment, such as the movement of water and the cycling of nutrients. But these models were not up to the task.

For one thing, they demand a lot of highly accurate input data. But there weren’t enough ground-based sensors to provide that data at the scale needed. Physics-based models also tend to carry a lot of uncertainty. And they are thirsty for computing power. “To run them at scale, you need a supercomputer,” says Huang.

Huang was at an impasse. But then he had an idea. How about using AI models, training them on a combination of ground-based sensors and satellite data, and then using those models to predict soil moisture from a satellite image?

“I thought AI could improve the predictive power of the data,” he says. “So I started to build a machine-learning model, which can see much deeper. With machine learning, you can understand much more about images.”

–-–

You might wonder how AI can detect soil textures and moisture from satellite images that stump a highly trained soil scientist.

Machine learning, which is a subset of AI, is a process by which a computer system learns from experience. As more and more data are fed into the system, the model refines its predictions. This is how large language models such as ChatGPT work — they’re trained on huge stores of data. Over time, as more and more users interact with the chatbot, the system gains accuracy and “intelligence.” Image-recognition algorithms, like the AI for detecting soil moisture levels, work similarly. After training, they can learn to identify subtle patterns in images that defeat the human eye.

Last year, together with collaborators at UW, the U.S. Department of Agriculture, and in Canada, Huang’s team got to work. They started building a machine-learning model to predict daily soil moisture content in crop fields across the U.S. at a resolution of 100 meters. The model would determine water content in soils at the surface and at the root zone. And it would update weekly, with the maps released to the public.

The team trained the model on thousands of satellite images of soil surfaces across the United States. Huang labeled those images with the measurements of soil moisture that his team had obtained through proximal soil sensors. Huang’s algorithm then learned to identify minuscule details in the images, down to the size of a single pixel, which it could then associate with soil moisture levels. Through repeated training, the AI came to recognize textures, patterns, and colors that correlated with different amounts of soil moisture.

“By combining multiple remote-sensing data sources and integrating them with machine-learning and data assimilation methods, the framework we are working with is achieving the desired reliability and accuracy,” says Shuohao Cai, a Ph.D. student in soil science who works with Huang. “Our soil moisture products will help farmers a lot in the near future.

–-–

From predicting the best times to apply pesticides to increasing carbon capture in the soil, AI is now solving problems across soil science. “AI is well suited for these problems because, compared to traditional modes, it has better predictive power of soil’s complex structure,” says Huang. “It’s also much faster and more efficient compared to process-based models. If you want a tool that can work at high resolution for the entire U.S., then you need AI.”

Huang’s initial results, published in February 2024 in Agronomy, were promising. This year, he released the code of the model to the public, allowing users from around the world to adjust the model for use outside the U.S. But Huang found the AI still struggled to predict water content below ground level.

“The machine-learning model was really good at predicting soil moisture at the surface,” he says. “But when we tried to predict moisture in the root zone, the model did a worse job.”

There was not enough high-quality training data to allow the model to reliably penetrate underground. So, Huang decided to combine the machine-learning model with elements of a traditional, physics-based model — a process that is ongoing. And Huang, in collaboration with colleagues in the lab of UW engineering professor Joseph Andrews, is also developing accurate low-cost sensors to gather even more data for his training models.

These sensors could be distributed across the U.S., harvesting data to improve the accuracy of the machine-learning tools. These sensors detect more than soil moisture. They also collect data about nitrate levels.

“One of the biggest issues we have in Wisconsin is a lack of reliable nitrate leaching data,” says Matt Ruark, associate professor of soil science who is working with Huang on the sensor project. “Most traditional approaches to quantifying nitrate leaching are laborintensive. Dr. Huang’s sensors allow us to answer the following questions: How much nitrate are we losing exactly? And does this vary by soil or cropping systems?”

All of this information will be fed into the AI. “We are going to combine this information with our traditional machine-learning algorithm and the domain knowledge that is included in a physics-based model,” says Huang. “In this way, we will overcome the gaps in the satellite data, and the model’s performance will improve. It will have the advantages of both machine-learning and physics-based models.”

Next, Huang plans to test the model by applying it to fields in Canada, where there is less training data. And Huang is looking forward to sharing the finished product with Wisconsin farmers over the coming years. “We are excited,” says Huang, “to distribute our forthcoming product to our farmer collaborators.”