Downscaling of seasonal precipitation for crop simulation

A non-homogeneous hidden Markov model (NHMM) is used to make stochastic simulations of March–August daily rainfall at 10 stations over the southeastern United States, 1923–98. Station-average observed daily rainfall is prescribed as an input to the NHMM, which is then used to disaggregate the rainfall in space. These rainfall simulations are then used as inputs to a CERES crop model for maize. Regionalaverage yields derived from the NHMM rainfall simulations are found to correlate very highly (r = 0.93) with those generated by the crop model from observed rainfall; station-wise correlations range between 0.44 and 0.74. Rainfall and crop simulations are then compared under increasing degrees of temporal smoothing applied to the regional-rainfall input to the NHMM, designed to exclude the sub-monthly weather details that would be unpredictable in seasonal climate forecasts. Regional yields are found to be remarkably insensitive to this temporal smoothing; even with 90-day lowpass filtered inputs to the NHMM, resulting yields are still correlated at 0.85 with the baseline simulation, while station-wise correlations range between 0.18 and 0.68. From these findings, we expect regional maize yields over the SE United States