Methods for Yield Forecast Based on Crop Model with Incomplete Weather Observations

Crop simulation models are important tools for identifying climate-crop relationships as well as for yield prediction, while complete daily weather data for whole growing season is required for running crop model, which cannot be satisfied only by weather observations in the real-time operation. Formal studies have generally used averages of daily weather calculated from the historical weather database as replacement, which may destroy its temporal distribution, and thus introduce another source of bias. Aiming at preparing meteorological data after the forecasting day that required by the crop model in the real-time yield forecasting operation, the climate analogues methodology is proposed, which can generate new climate series for the desired period from history observations that with similar climates across space and time, based on a distance metric such as Euclidean, and the new proposed methodology is tested with the CERES-Rice model for its predictability and error distribution, comparing with a general arithmetic mean method. Results show that rice yield is sensitivity to meteorological conditions during two months before maturity, yield forecasting with CERES-Rice model driven by weather data at two-month lead-time leads to a more than 60% prediction probability with an error no more than 5%, and such predictability increases steadily with weather observations updated, showing considerable potential for operational application. Considering there is no priori knowledge on the climate trend for the remainder growing season, using a multi-year mean weather data instead, there is a 60% prediction probability when forecasted at two months before maturity and a 70% prediction probability one month before, however, obvious systematic overestimate is observed, and there exist systematic errors among different decades using 30-year means due to the climate trend under global warning, by using the latest 10-year or 5-year means, the decadal systematic errors decrease while the predictability increase for the poor ability in representing climate variability among years. Finally, using the historical analogue approach that generating downscaled daily weather data from historical observations, the prediction probabilities increase slightly, while the systematic errors reduce considerably compared with that of using the general arithmetic average approach, in addition, the historical analogues approach allows to include climate trend for the upcoming growing season, and by doing so, the predictability increases to more than 80% at two month in advance, much higher than that with multi-year mean. It is concluded that the analogue approach has great potential in bridging the gap between crop model and climate forecasting.