Introduction

Gro’s Yield Forecast Models use a suite of machine-learning models to estimate in-season yields at the district, province, and/or national levels on a daily basis. The US Soy Yield Forecast Model updates daily and predicts the difference between the estimates in USDA’s June acreage report and final yield figures. 

Customers Use the Model to:

  • Predict future trends such as crop yields and production, agricultural input demand, and significant weather patterns
  • Gauge crop availability and crop prices
  • Understand how weather impacts yields in microclimates
  • Inform other models focused on damage caused by pests and diseases
  • Monitor crop production, the most significant part of the balance sheet, by tracking its most variable component, which is in-season yield
  • Obtain early, district-level estimates of the final area of soybeans planted

Why It Matters

Soybeans are significant as an input in the livestock industry, as a raw material for biofuel, and for human consumption. The US is the second-largest producer and exporter of soybeans, and changes to its crop have ramifications on a global scale, both for other key soy producers and for livestock producers. As such, following estimates of US production throughout the season can guide decision-makers across the value chain of the crop.

Methodology

We modeled county-level yields and then aggregated them to the national level, weighting by reported harvested area. The input variables of the model included: 

  • Daily land surface temperature (LST)
  • Eight-day time-step normalized difference vegetation index (NDVI) from moderate-resolution imaging spectroradiometer (MODIS) satellite images
  • Gridded daily soil moisture estimates from satellite images (SMOS)
  • USDA-reported crop conditions
  • Percentage of soybean harvested area that was irrigated (assumed constant across years)
  • Static soil condition data
  • Elevation data
  • Crop calendars
  • Latitude and longitude of each county

 

Next, we aggregated the gridded spatial variables to the county level. We applied our soybean land cover to soil texture data, elevation, and the percentage of irrigated fields. We evaluated our model performance through ”leave-one-year-out” cross-validation using 2009-2016 data.

Related Resources