Forecasting Cotton Yield in the Southeastern United States using Coupled Global Circulation Models

We developed methods of forecasting cotton (Gossypium hirsutum L. var. hirsutum) yields at a county level 3 mo before harvest for the states of Alabama and Georgia. Cotton yield historical records for 57 counties were obtained from NASS and detrended using a low-pass spectral filter. A Canonical Correlation Analysis regression-based model was annually recalibrated to incorporate the year-by-year accumulating data: (i) April–June (during vegetative growth) observed rainfall for the forecasting year, and (ii) July–September (during reproductive growth) global-scaled 2-m mean temperatures for years before the forecasting year, beginning with 1970. We produced two types of forecasts: short range and medium range. The short-range near-term yield forecast (just before initiating harvest in the region) used gridded assimilated observed 2-m mean temperatures obtained from the NCEP-NCAR CDAS Reanalysis data. The medium-range forecast (3 mo before harvest) used 2-m mean temperature retrospective forecasts from the operational NOAA/NWS/NCEP Climate Forecasts System coupled global circulation model. The short-range, near-term forecast performance was measured by leave-one-out cross-validation and retroactive validation, whereas medium-range forecast performance used the previous two methods plus a proposed coral-reef validation method. The agreement between short-range near-term forecast and actual cotton yield was statistically significant at the 0.05 level in 31 out of 57 counties. For 48% of these 31 counties, the agreements between medium-range forecasts and actual cotton yields were statistically significant at the 0.05 level. The goodness-of-fit index for those 15 counties was 0.51 and the RMSE ranged from 13 to 31% of the annual yield.