By Kimberly Mann Bruch, ACCESS, and Ken Chiacchia, PSC
Maryland Researchers Advance Agriculture Dashboard Using ACCESS Resources
Every season, farmers face an age-old problem: How can they plant, fertilize, and water their crops for the best possible yield? A team led from the University of Maryland has used two supercomputers — PSC’s Bridges-2 and Johns Hopkins University’s Rockfish — to create and run DAWN, an artificial intelligence (AI)-enhanced, interactive program allowing farmers to explore different crop management options given the weather expected from national climate predictions.
WHY IT’S IMPORTANT
Farming is one of the most important, and hardest, jobs on the planet. Our modern world would come crashing to a halt if we all couldn’t depend on people who specialize in growing food for the rest of us.
But nature doesn’t make farming easy. From the beginning of agriculture about 10,000 years ago, farmers have used every trick they could think of to decide what crops to plant, when to plant, and how to plant. Making the right decisions — growing the right food crops at the right time, with the right fertilization and water management — has long made the difference between plenty and starvation. It can be a little bit better today, when sophisticated meteorology makes the weather less of a question mark, and global trade can help make up shortfalls locally. But as we know, supply-chain problems can make it hard to transport food, and in any case, a bad year can put farmers out of business and cripple our ability to grow food next year.
“DAWN aims to empower stakeholders with the ability to access and utilize its capabilities to enhance land, water, and fertilizer management across various agricultural systems and scales … Our dashboard tools are co-produced with farmers so that they are tailored to their needs and effective for agricultural decision making.”
— Xin-Zhong Liang, UMD
A team led by Professor of Atmospheric Science Xin-Zhong Liang at the University of Maryland (UMD) decided that modern farmers could succeed better if they had a computerized “dashboard” that integrates different types of weather and climate prediction — such as the National Oceanic and Atmospheric Administration’s operational seasonal climate forecasts — in a simple tool that helps them make growing decisions. To build their tool, Dashboard for Agricultural Water Use and Nutrient Management (DAWN), they turned to two supercomputers in the NSF ACCESS system — Rockfish at Johns Hopkins University and Bridges-2 at PSC, a leading member of ACCESS.
HOW PSC HELPED
DAWN would need to couple climate-crop simulations with emerging AI technology. Assistant research scientist Chao Sun, a member of Liang’s team, aimed to provide farmers with multidisciplinary insight on their specific farming area’s climate and crop prediction with a six-month lead time. He and Liang’s team worked with farmers to ensure that DAWN’s interface would offer them the information they need in a user-friendly way.
The scientists’ dashboard tool would start with regional high-resolution climate-crop coupled numerical model predictions for the entire continental U.S. as well as the Gulf of Mexico. By pairing these data with complex AI algorithms, they could improve on the predictive power of climate-only forecasts. DAWN’s AI algorithm would look at past crop progress and compare it with predictive future situations based on precipitation, temperature, irrigation, fertilizer, and more.
“The complex algorithms and large memory required to run DAWN wouldn’t have been possible on our local machines; hence, we reached out to the NSF for ACCESS allocations … Bridges-2 and Rockfish provided us with the power needed for our work.”
— Chao Sun, UMD
DAWN presented a computational challenge, in that it required computers that could both handle massive data and had the AI-friendly graphics processing unit (GPU) nodes that accelerate the learning process, in which an AI first trains itself on labeled data so that it can reliably make predictions on data that humans haven’t labeled. PSC’s flagship Bridges-2, for example, was perfect for the job, offering more than a thousand fast central processing units (CPUs) optimized for quick calculations and fast data movement, as well as 280 late-model GPUs that are best for the comparison tasks needed for AI learning.
The UMD scientists published their results in the journal Bulletin of the American Meteorological Society. You can read more about the DAWN project here.
Now that DAWN is ready for use, the UMD team is working with early users to investigate how farmers can best employ the tool and identify ways to enhance the dashboard with additional tools and features with planned subsequent updates. One promising feature of DAWN is its ability to help farmers explore different crop management options to see how they affect yields in the long run. With their partners in the farming community, the team is also assessing the limitations of the system, with an eye toward addressing them and making DAWN even more useful. The scientists have proposed a series of future experiments, running the software every five days for a year, to assess seasonal crop predictions more intensively.