Interannual variability characteristics and simulated crop response of four stochastic weather generators

Abstract

Many stochastic weather generators commonly used with crop models tend to under predict interannual variability of climate and, as a result, distort distributions of crop simulation results. We examine the ability of four stochastic weather generators, WeatherMan, MARKSIM, WM2 and LARS-WG, to reproduce interannual variability of monthly climate and crop simulation results. Comparisons were based on bias and RMSE of means and standard deviations of monthly precipitation totals, frequencies of wet days, mean daily temperature extremes for 12 long-term weather data sets; and yields and dates of anthesis and harvest maturity of three crops under a total of 10 scenarios simulated with dynamic crop models. Evaluation also considered statistical tests of equality of distributions of the same variables between observed and generated weather data sets. The generators generally reproduced climatic means well, except for MARKSIM which showed positive bias of mean monthly rainfall totals and frequencies. WM and LARS-WG showed substantial negative bias of interannual variability of monthly precipitation totals. As a result of stochastic resampling of wet-day frequencies, MARKSIM and WM2 showed little negative variability bias of monthly precipitation totals, but positive variability bias of monthly wet-day frequencies. WeatherMan, MARKSIM and LARS-WG under represented variability of monthly mean temperatures, whereas WM2 showed positive but smaller mean temperature variability bias. WM2 reproduced simulated distributions of yield and harvest maturity dates better than the other generators. Results for simulated anthesis dates were not consistent among generators. This study supports the need for some form of low-frequency variability correction in stochastic weather generators for applications in which reproducing interannual variability is important. Results generally favor WM2 over the other generators tested for applications in which variability of simulated yields is of primary importance.