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. Gro’s Canada Spring Wheat Yield Forecast model uses inputs including temperature, rainfall, and soil moisture. The model provides daily, in-season yield forecasts at the district level across Canada’s largest wheat-producing provinces. The estimates are then aggregated to the provincial and national levels.

Customers Use the Model to:

  • Gauge crop availability and crop prices
  • Predict future trends such as crop yields and production, agricultural input demand, and significant weather patterns
  • Understand how weather impacts yields in microclimates
  • Inform other models focused on damage caused by pests and diseases, as well as for price forecasting
  • Monitor crop production, the most significant part of the balance sheet, by tracking its most variable component, which is in-season yield
  • Monitor and predict Canada spring wheat yield at the Small Area Data region level. 

Why It Matters

Canada is the world’s sixth-largest producer of wheat, and the fourth-largest exporter of the grain. Spring wheat represents roughly three-quarters of Canada’s total wheat production. The high-protein variety produced by Canada serves a different function in global markets than the more common lower-protein variety found in most of the US, EU, Black Sea, and Argentina. The crop’s significance to national and global economics makes it important to track yield figures in the country.

Methodology

As with all of Gro’s yield models, our Canada Spring Wheat Model updates daily at the district level. The machine-learning model takes into account a multitude of datasets and uses 77 environmental, climate, supply, and acreage-related variables to generate the best daily forecasts throughout the growing season. 

Gro’s model has proved to be more accurate than in-season yield estimates provided by StatCan, Canada’s national statistics office. Over a decade of backtesting, the Gro model has had a mean absolute percentage error (MAPE) of 4.1% for forecasts made on Sept. 1 that aim to predict final yield numbers published in December. By contrast, StatCan’s survey-based yield estimates in September have had an 8.6% error rate over a similar time period.

The model runs from April to October, and uses the following inputs: 

  • Location data
  • Gro crop-weighted NDVI (normalized difference vegetation index)
  • SMOS
  • Evapotranspiration
  • Land surface temperature (LST)
  • Soil grids

Related Resources