Crop Production Risk in the Pampas: A Bayesian Weather Generator for Climate Change and Land Use Impact Studies

Abstract:

We present a space-time stochastic weather generator for daily precipitation and temperature, developed within a Bayesian hierarchical framework (hereafter BayGEN). This framework offers a unique advantage: it provides robust estimation of uncertainty that is typically under-represented in traditional weather generators. Realistic estimates of uncertainty are of utmost importance for studying climate variability and change, impacts on land use, and crop production. BayGEN is applied to a network of weather stations in the Salado basin of the Argentine Pampas, a region that saw immense agricultural expansion towards climatically marginal (i.e., semi-arid) regions, in part due to significant trends in annual precipitation from 1970-2000. Since the turn of the century, observed conditions suggest a decrease in precipitation, which begs the question: "Are the existing agricultural production systems viable in a drier future?" The use of process based (i.e., hydrologic, crop simulation) models in conjunction with BayGEN will allow for complete analysis of the system's response to an ensemble of plausible future scenarios. Precipitation occurrence at each site is modeled at the first level of hierarchy using probit regression with covariates for seasonality, where the latent process is Gaussian -- positivity in the latent process implies occurrence. The precipitation amounts are modeled using a transformed gamma regression (i.e., gamma generalized linear model), similarly with seasonality covariates. Minimum and maximum temperatures are conditional on precipitation occurrence and are decomposed into two processes: (i) climate -- linear regressions on seasonality covariates, and (ii) weather -- realizations from mean-zero Gaussian random fields. The use of seasonality covariates allows the generation of daily weather sequences conditioned on seasonal forecasts or projected multi-annual trends, an increasingly important practice for risk assessment in climatically marginal regions. Posterior distributions of the regression and spatial parameters are used to conditionally simulate daily weather at observed locations and also on a regularly spaced grid (i.e., unobserved locations) for use in agriculture and hydrologic modeling.