Issue link: http://read.uberflip.com/i/1443765
34 POTATO GROWER | FEBRUARY 2022 DIGGIN' IN WATER | By Sean W. Fleming, USDA-NRCS New river forecast model integrates AI to support water management in the West Quenching Your Thirst Water supply forecasts are important for any crop year. But for farmers, ranchers, foresters and water managers in the West facing extreme and debilitating drought conditions, those forecasts have never been more critical to their operations and livelihoods. Since the Dust Bowl of the 1930s, USDA's Natural Resources Conservation Service (NRCS) has helped America's producers plan for their operations through the Snow Survey and Water Supply Forecast program. The program runs a massive network of mountain climate and snow- monitoring sites across the western U.S. called SNOTEL. This is coupled with other data and computer models to predict the amount of river runoff in the upcoming spring and summer. These water supply forecasts are used by America's producers to plan their operations for the year by helping guide choices like crop selection, water rights rentals and whether to leave land fallow. Over the decades, that information has grown to be used by many other groups for many purposes—from optimizing hydroelectric power generation, to assessing seasonal flood risk, to complying with legal decisions around endangered species and international treaties governing transboundary rivers. The value of water managed using these forecasts is easily in the billions of dollars, and even modest increases in accuracy can create over $100 million a year in public benefit for just one river basin. However, major forecasting improvements are needed because of narrowing margins between water supply and water demand in the ever-more-thirsty American West. Those tighter margins reflect a combination of climate change and population growth, and they mean there's less room for error than ever before in water management, requiring improved efficiency and accuracy in everything we do. NRCS has unveiled a new computer application to address this pressing need: the multi-model machine learning metasystem, or M4. This first- of-its-kind model will be the largest migration of artificial intelligence (AI) into real-world river prediction programs. Researchers first experimented with machine learning, a branch of AI, for hydrologic forecasting a quarter-century ago. But they couldn't jump the research-to- applications gap—the needed step of getting from what works in the lab to what works in the field. Ironically, scientists and engineers working outside the tech sector have often been the last to adopt AI into their everyday practices. Unlike some other areas, STEM fields have long used sophisticated math and computer models. AI had to successfully compete with those existing methods to gain widespread acceptance, which in many fields, including earth and environmental science, is only starting to happen now. The average hydrologist is still more likely to use AI—in a smartphone app, for example—to find the quickest route to the office in the morning, than to apply it in their work when they get there. The M4 project aimed to change that. Applied scientists at NRCS took a pragmatic approach: They looked in detail at what they needed in the next generation of their operational river forecast system, and then created a new tailor-made solution from existing building blocks. That included adopting automated machine learning, which makes it easier and faster to use, and radically improving the explainability of the results, putting to bed a longstanding worry about so- called "black box" AI technologies. Testing proves the system is more accurate, robust and simple-to-