What do you get when you combine seasonal climate prediction, crop growth information, satellite imagery and apply sufficiently powerful HPC? A sophisticated maize crop yield prediction tool that may prove more accurate than the USDA model that is the current gold standard.
The key differentiator in the new approach, developed by Blue Waters Professor Kaiyu Guan and National Center for Supercomputing Applications postdoc fellow Bin Peng and detailed in the journal Geophysical Research Letters, is the inclusion of both climate and remote sensing observations.
“Compared with using historical climate information for the unknown future, which is what most previous research is based on, using seasonal climate prediction from the NOAA’s National Centers for Environmental Prediction gave better forecasting performance, especially in reducing the uncertainties,” says Peng, the lead author of this study.
Guan, a co-author, adds, “But if we only use seasonal climate prediction data – temperature, rainfall, and vapor pressure deficit – our predictions were no better than the USDA’s. It was only when we added the satellite data that we started to see the improvement. That’s a clear indication that satellite data is extremely useful in this case.”
Accurate crop yield predictions are an essential part of the U.S. economy from the regional level up through the largest enterprises. Governmental agencies and private businesses alike rely primarily on the U.S. Department of Agriculture’s monthly World Agricultural Supply and Demand Estimates (WASDE) reports to get their crop forecasts. This study, powered by the Blue Waters supercomputer at NCSA, found that WASDE reporting was less accurate than the University of Illinois’ “climate+remote sensing” model in predicting end-of-season yield. As an example, on average, the WASDE report for June between 2010-2016 was off by 17.66 bushels per acre, while Guan and Peng’s system improved that delta to 12.75 bushels per acre. For the same time period, WASDE’s August predictions were off by an average of 5.63 bushels per acre, while the University of Illinois research reduced the number to 4.37.
The research was supported by NASA, USDA NIFA, and NCSA, which receives support from both the NSF and the state of Illinois for its Blue Waters supercomputer.