There are two crop seasons for wheat in the US. In Northern states such as North Dakota, South Dakota, and Montana the wheat crop is mostly planted in the spring and harvested in the fall. In some warmer Southern states, such as Kansas, Oklahoma, and Texas, wheat is planted in the fall and harvested the following summer. In addition to the crop seasons, wheat crops are further classified in the US as hard red, soft red, hard white, and soft white. This is done for a number of reasons including protein content and trading purposes. Because these wheat varieties grow in different locations and seasons, and differ significantly in yields and suitable weather conditions, we need to build separate yield models for each variety.

Our US hard red winter wheat model is built at the county level for three states—Kansas, Oklahoma, and Texas—which together account for about 60% of total production of hard red winter wheat (referred to as HRWW) of the US. Because the USDA doesn’t provide county-level yield of HRWW, we used winter wheat yield to train our model against, which is a good approximation to HRWW yield in our study area, since they grow HRWW almost exclusively.

Among all the geospatial inputs we tested for our model, normalized difference vegetation index (NDVI) and evapotranspiration (ET) showed best performance during the backtest. Besides the geospatial inputs, weekly winter wheat crop conditions from the USDA (at the state level) also helped improve the model’s performance. By the end of June, 72% of counties were predicted within 8.14 bushels per acre.