Stochastic disaggregation of monthly rainfall data for crop simulation studies

Abstract

Stochastic weather generators are useful for producing daily sequences that reproduce climatic statistics aggregated to, e.g., a monthly time scale, for use with biological simulation models. This paper describes a stochastic weather generator that disaggregates monthly rainfall by adjusting input parameters or by constraining output to match target rainfall totals, and demonstrates its use with a maize crop simulation model at three locations. Constraining generated monthly rainfall to match observations reproduced the cross-correlation between observed amount, frequency and mean intensity of rainfall more nearly than conditioning weather generator parameters on monthly rainfall. The use of stochastically disaggregated observed monthly rainfall accounted for between 55 and 93% of the variance of yields simulated with observed daily rainfall. Constraining weather generator output required roughly an order of magnitude fewer realizations than conditioning input parameters, to achieve 95 or 99% of the asymptotic maximum correlation with yields simulated with observed rainfall. Negative mean bias of yields simulated with rainfall disaggregated by adjusting weather generator parameters may be due to the variance among stochastic realizations and nonlinearity of yield response to water availability. For maize yields simulated with GCM rainfall predictions at Katumani, conditioning frequency parameters on predicted monthly totals gave the lowest systematic and random error with 1000 realizations. However, constraining generated rainfall to monthly predictions gave better results with few realizations. Although most of the predictability of monthly rainfall was associated with predictability of rainfall frequency, there was little benefit from incorporating predicted frequency into stochastic disaggregation.